Changelog

What's new in Agentery — the AI agent & MCP pricing-intelligence engine. Shipped continuously.

5 Nov 2026

Price checks judge each offer on its own evidence

Our nightly checks no longer set aside a whole product's price just because its page has switched between two figures several times. A price is held back only when the page shows evidence for that specific offer, such as both figures printed together or conflicting billing labels.
A price that rises and later falls can now count as two real changes, and genuine promotions are no longer excluded by default.
Amounts printed only as a saving ("Save $100") or as a comparison ("under $500") are no longer read as a plan's price.
Tonight's scan stopped partway when its browser restarted; it was resumed and the 5 November AEPI release published at 05:21 UTC.

4 Nov 2026

Listings now say when a price may be out of date

When a pricing page has changed since we last read its price, the listing page and the provider profile in our MCP now say so, with the date the change was seen. The price stays shown until it is read again.
The daily AEPI release now reports how many of its listings have an up-to-date price reading.

4 Nov 2026

Today's price update was late

The nightly price scan stopped at 02:02 UTC when its browser ran out of memory. It was restarted and today's AEPI (release 173, 4 November) was published at 08:23 UTC, about six hours later than usual.
The scan now restarts its browser regularly and recovers by itself if it stops, with an alert if that fails.

3 Nov 2026

"MCP endpoint" links on listing pages work again

A vendor told us the "MCP endpoint" button on their listing opened a broken page (/agents/[object Object]).
The endpoint addresses were stored correctly, but the listing page displayed them wrongly. This affected every listing with a published MCP, API or A2A endpoint, and is now fixed for all of them.

3 Nov 2026

Very small per-use prices now show in full

A vendor told us their $0.002-per-credit price still showed as $0.00 on the plan card, and as "$0.002/mo" in the value summary.
Prices below one cent now show their full digits everywhere on a listing page, and per-use prices show their unit (for example "/credit") rather than "/mo" or a bare slash. This applies to every listing: about 1,100 have sub-cent prices and 969 had per-use prices with no unit recorded.
Display only: the price index and stored prices are unchanged.

2 Nov 2026

Plans marked "coming soon" are no longer shown as current prices

A vendor told us their listing showed $9.13 a month because two plans on their page, labelled "Coming Soon — Not available to purchase", were read as current subscriptions. They were not for sale yet.
We checked every listing for the same problem and found 104: plans marked coming soon, waitlist-only, "not on sale yet", or a pre-order you only pay for when it ships. 102 listing pages were corrected today; on 52 of them the lowest price shown changed. One more is left for review because its page changed overnight.
These plans are now kept as planned plans, with their price and wording, and shown on the listing as "Planned, not on sale yet". They no longer count as the listing's price, in price comparisons or in benchmarks. The daily price reading applies the same rule from now on, and a plan returns as soon as the vendor puts it on sale.
Wording that is about something else does not remove a price: a feature marked coming soon inside a plan you can buy, or "annual billing coming soon" beside a monthly price.
AEPI: published releases are unchanged. From the next nightly reading these plans leave the index through its normal checks.

1 Nov 2026

AEPI currency fixes: the vendor's own price is the purchase price (methodology revision 1.6-fx1)

When a plan card shows the vendor's own price next to a converted figure in another currency (for example '$10.26 per month EUR 9/mo', 'EUR 43/mo ≈ $49/mo' or 'EUR 21 /month (RM99/month)'), the index now uses the vendor's own price and its currency. Before, the converted figure was sometimes read instead, and the plan then dropped out of the index. The two figures must agree at the ECB rate of that day; the converted one is recognised by an approximate sign, or by odd cents beside a whole price.
One currency written in different ways ('$', 'US$' and 'USD'; 'EUR' and the euro sign; 'GBP' and the pound sign) now counts as one currency. A dollar price printed with a code beside it ('CAD $29', '$50 CAD') is in that currency. Malaysian ringgit ('RM') is recognised.
The whole history from 30 Jun was rebuilt with these fixes (release 164); the definition of the index did not change and earlier releases stay published. Effect on the 1 Nov headline: 100.212 -> 100.219 in US dollars. 88 listings changed: 83 gained days in the index, 3 lost some, and 15 now carry a different or newly recognised currency.
Known gaps stay gaps: the GNK token has no ECB rate, so no exchange-rate factor is applied to it, and nothing is filled in.

1 Nov 2026

AEPI data: currency attribution, revision dates and data gaps now in every index response

The price index tools (MCP get_price_index and get_price_index_history, and the REST index endpoints) now return the reference currency (US dollars), the methodology revision with its own date (the index v3 was first published on 24 Oct 2026; methodology 1.6 on 31 Oct 2026) and, in full mode, the exact split of each change into the quoted-price part and the exchange-rate part, read from the published release. Log changes of the two parts add up; index-level factors multiply.
Full responses also show how the currency of the index weight is known: printed currency code or unambiguous symbol, a stated rule, or a bare '$' counted as US dollars. About 64% of the weight on 31 Oct is a bare '$' counted as US dollars, so currency uncertainty is not limited to the small unresolved share.
The daily count is now named for what it is: the number of priced purchases compared that day (quoted_price_comparison_count). The old field like_for_like_pair_count stays as an alias; it is not a count of strict like-for-like pairs.
The history lists days without a published observation, with the reason. 29 Jul 2026 has none: that night's full page captures were lost in a server incident, so the day is not rebuilt and no value is filled in.
Price changes of 5% or more (price after / price before - 1, in the quoted currency) are highlighted in the movers lists and the daily commentary. Smaller supported changes still count in the index and its history.

1 Nov 2026

Price changes behind the index: latest release only, product and % change

Wherever we list the product price changes behind an index, sub-index or sector move (the AEPI page, index and sub-index charts, sector pages, the homepage "Where prices changed" box, the API and the MCP tools), we now show which products moved and by how much, for the latest release only.
Each change is shown as the product (and plan) with its percentage change, rounded to a whole percent, and its direction. Old and new price amounts and the dates of individual changes are no longer listed, and the day-by-day and 7-, 30- and 90-day lists of earlier changes have been removed.
Index levels, index history and charts are unchanged, and every supported price change still counts in the index. Changes of 5% or more are still highlighted. Nothing has been deleted.
Daily market commentary now names products with their percentage change only. Posts already published are unchanged.

31 Oct 2026

AEPI methodology 1.6: the index now follows publicly quoted prices, in US dollars

After a human review of 19 price cases, the Agent Economy Price Index (AEPI) now measures changes in the publicly quoted price of the same product and offer. Advertised starting prices, eligible promotional prices and the end of a founder rate now count, each labelled with its role. A struck-through list price is never treated as payable. An introductory first-months price is kept separate from the renewal price.
When a plan's price changes together with what it includes (credits, events), the price change counts and the change in what is included is recorded beside it. The stricter like-for-like judgement is still kept on every change.
Prices quoted in other currencies now count in US dollars, using European Central Bank daily reference rates. Stablecoins use their observed daily rate. Each such change is split into the vendor's own price change and the exchange-rate effect. A converted price added next to an unchanged one is not a price change, and a dollar price is never converted twice.
Monthly/yearly switches are settled from the page's own billing labels. A ratio that happens to match a yearly discount is never treated as proof either way. Documentation examples and pages that are not product offers are not prices.
The whole history since 30 June was rebuilt under these rules, and the headline, sub-indices, sectors, API and MCP answers come from one new release. Earlier releases stay published. Moves of about 5% or more are highlighted in commentary; smaller changes still count in the index.

31 Oct 2026

25 more directories feeding new listings, and four sources stopped because their terms forbid it

We checked 100 directories and marketplaces of MCP servers, AI agents, APIs and AI tools against our sources. 25 that we were not using are now read twice a day, each under its own source tag: MCP Market, Agent Connectors, The AI Agent Index, AgenticList, aiagents.wiki, the Gemini CLI extensions gallery, ClawHub (OpenClaw plugins), MCP.Directory, MCPServers.com, MCPDirectory.dev, PopularAiTools.ai, RankmyAI, AI工具集 (ai-bot.cn), Dang.ai, AI Tools Directory, PoweredbyAI, AIToolly, AITopTools, APIs.guru, bai.tools, Toolspedia, DokeyAI, AI Center, ToolAI and AIParabellum.
We only read a directory when both its terms and its robots.txt allow automated reading. We use its official API, feed or sitemap where it has one, and we list the product's own website or code repository, never the directory page. Every candidate still passes the usual screen before it is listed.
Generic AI-tool directories list mostly image, video and writing tools. From those we only take entries that read as AI agents, MCP servers or agent tooling.
The first pass queued 389 new candidates after removing products we already list.
We stopped four sources whose terms forbid automated collection: Product Hunt, allagents.app, AIxploria and powerusers.ai. Products already listed from them stay listed.
58 other directories were not added: their terms or robots.txt forbid automated reading, they block automated visitors, or they list no outside products. This includes G2, Capterra, GetApp, AlternativeTo, Zapier, Make, RapidAPI, Postman, OpenRouter and Composio. Agentery was not submitted to any directory.

31 Oct 2026

Agentery MCP: the first answer now says what it is based on and how to sharpen it

Every task answer (search, compare, price benchmark, market report, research and alternatives) now opens with a short answer_basis block: the unit, buyer tier, billing period, currency, product type and how many products the figures cover.
It also lists what you told us and whether each point was applied, what the tool chose because you did not say (for example all buyer tiers), which missing details would change the answer, and the exact arguments for a sharper follow-up call.
We no longer guess your team size. When products are priced per seat and you have not given a seat count, the answer shows the cost at 1, 5 and 20 seats from each vendor's published per-seat price and seat minimum, clearly labelled as scenarios. If you give a seat count, the cost is worked out at that number.
Clearly meaningless queries (such as "12345" or "test") get a short request for a job description with an example, instead of a search.
A follow-up sent with an expired or unknown task reference is now answered from its own question, with a note saying what happened to the reference.
Prices and rankings are unchanged; the new block sits beside the existing answer. MCP version 1.27.0.

31 Oct 2026

More products from China, India, Russia, Brazil, Germany, France, Spain and Latin America, Turkey, Indonesia and Vietnam

Agentery now looks for AI agents, MCP servers and agent tools in ten more markets, in the places their developers publish: GitHub projects described in Chinese, Russian, Portuguese, German, French, Spanish, Turkish, Indonesian, Vietnamese or Indian languages; GitHub projects built for local services (for example Indian stock exchanges and UPI, Brazilian Pix and NF-e, German legal codes, Turkish court and statistics data, Vietnamese shop systems and Zalo); community lists of local MCP servers; and makers' launch posts on V2EX (China) and TabNews (Brazil).
Every find goes through the same checks as any other listing: duplicates are dropped, each candidate is screened, and only genuine products are published. The first run added 98 listings across these markets.
Listing text is written in English so it can be searched and compared. Product names keep their original script, with an English or romanised form stored for search.
A listing shows a country only when there is real evidence for it, such as a registered address or the maker's stated location. The language a product is written in never sets its country.
We only use sources whose terms and robots rules allow this kind of reading. Sources that forbid it are not used.

31 Oct 2026

Three new ways we find products to list

Products mentioned on pages we already track: when a listed product's site names another product (integrations, "works with", alternatives, comparisons, partners, related projects), we now check that product and, if it is a live AI agent, MCP server or agent tool we don't list yet, it goes through the same review as every other new listing.
Business software marketplaces: we now read the AI and AI-agent sections of the public catalogues whose terms and robots rules allow automated reading (Microsoft AppSource/Azure Marketplace, the Slack App Directory and Google Cloud Marketplace). Marketplaces that don't allow it are not used at all.
Newly registered sites: we sample public Certificate Transparency logs for newly secured web addresses that mention agents or MCP, and review the ones that are live product sites (parked pages, test and staging addresses are skipped).
Every new listing still has to pass the same checks: it must be a real, live product, it is not added twice, and its card is written in English (keeping the original name).

31 Oct 2026

Sub-indices now show the price changes behind them

Open any sub-index (Agents, MCP servers, Infrastructure, APIs, Platforms, Frameworks, Skills & plugins; the billing models Subscription, Hybrid, One-off and Pay as you go; A2A and the payment methods; or a buyer tier) and you now see the published price changes in it, as sector charts already do: product, plan, old and new price, percentage change, date, and whether the change pushed the index up or down. Each row links to the product.
Choose 7, 30 or 90 days (30 by default; on the index charts page the choice follows the chart's period). Newest changes come first. If nothing changed in the window, the page says so.
The list includes only price changes counted in the published index release, sorted into sub-indices the same way the index is. Currency-only moves are not listed, and how much each change moved the index is not shown. The same list is on /aepi for each scope.

31 Oct 2026

Price index: real price changes can now count on the night they appear

When two separate readings of the same two nightly captures agree that a plan's price changed, from the same old amount to the same new amount, and automatic checks confirm it, the change now counts in the index that night instead of waiting.
The checks are strict: both prices must appear word for word in the full saved pages, for the same plan, in the same currency and with the same billing period and wording on both nights. Price ranges, free-to-paid switches, promotions and figures that only reflect a monthly/yearly switch are left out. If anything disagrees, the change is not counted and is kept on a list for review.
Tonight (31 Oct) three changes were confirmed this way: Bankbridge (1 bank: $5 to $8 a month, and $50 to $80 a year) and Slid Phi Labs Chamber ($9 to $20 a month). The index moved from 100.54 to 100.57. The 30 other changes flagged tonight did not pass the checks and were not counted. Earlier history is unchanged.

30 Oct 2026

Ten products restored that had been removed by mistake

Ada, ManyChat, Engati, HappyFox, LiveKit, ZenRows, GTmetrix, Happy Scribe, CryptoQuant and Octane AI are listed again. They had been removed from the directory in error; nothing was wrong with the products. Their prices are refreshed in the next nightly check.

30 Oct 2026

Faster answers from the MCP price tools

The Agent Economy Price Index for one delivery type (for example agents only, or MCP servers only) could take 45 to 50 seconds when its cache was cold. It now answers in about a tenth of a second. The numbers are unchanged.
Asking about the same job again, in any of the market and price tools, is now quicker (under a third of a second in our tests instead of about a second), and returns exactly the same members each time.
A rare stuck step could make a market lookup wait until the 100-second limit. Each step that checks whether products do your job now gives up a few seconds after its own time limit, and those products are counted as 'unchecked' in the answer, as before.
A brand-new question still takes a few seconds, because each nearby product is checked against your job. We did not change that check, so the answers stay the same.
We checked the change by replaying 30 recent real requests before and after: none of the answers changed because of it.

30 Oct 2026

Salesforce Agentforce is listed again; hidden listings no longer dead-end a lookup

Salesforce Agentforce is back as a live listing in the customer-support AI agent market, linked to its official product and pricing pages. Its current pricing is read from tonight's scan.
It had been hidden in June by an automated clean-up of listings whose pages our reader could not open; that clean-up recorded no reason. We reviewed other well-known products hidden the same way and restored or re-linked the clear mistakes.
Looking up a product by an old id that was merged into another listing now returns the listing it was merged into. If a name only matches hidden listings, or only similar names, the answer says so and the name is sent for research together with the job you described.
This came from a real request: a buyer researching customer-support agents asked for alternatives to Agentforce and was told it did not exist.

30 Oct 2026

Same-name products: lookups now show every listing with that name

When you look up a product by a plain name that several listings share, Agentery still returns the listing with that exact id, but now also lists the other products carrying the same name, with what each one does and its website, ordered by how well they fit the job you described.
If the matched listing does not look like the product for your job, the answer says so, and the name is sent for automated research so the product you meant can be added. This applies to alternatives, provider profiles and comparisons.
This came from a real request: a buyer researching meeting transcription asked for alternatives to Fathom and was given an unrelated coding plugin that shares the name. Fathom, the AI meeting notetaker (fathom.video), is now listed.

30 Oct 2026

AEPI methodology 1.5.1: twelve wrongly posted price changes corrected, more products counted

The AEPI history has been recalculated under methodology 1.5.1 (release 153). Twelve price changes that had been posted wrongly were traced to their causes and fixed with general rules, not by excluding providers. Nine further false moves in the previous series are also gone: currency-conversion drift on pages showing several currencies, a struck-through reference price, and a page switching between prices with and without VAT.
Two of the twelve turned out to be genuine price cuts and are now posted correctly. A price rise previously excluded in review was confirmed from the stored pricing page and restored.
Products now join the index on any plan whose price, unit and billing commitment are clearly shown, even when other parts of the page are unclear; about 28% more products are counted. Prices are never read from a shortened copy of a page when the full capture is stored.
The headline for 30 October is 100.54 (previously 100.44). Earlier versions remain published and downloadable. Every new posting in this release was checked against its stored pricing pages by a model reading, not a human audit; detection of real changes is still incomplete and is the next piece of work.

30 Oct 2026

Listings now carry a country, and Japanese and Korean products are discovered directly

Agentery now records the country each vendor is based in. The country is read from evidence: a company address, registration or tax number on the vendor's own pages, a country website domain that requires a local presence, or the location on the publisher's GitHub profile. It is filled only when that evidence is clear, about 10,300 listings so far, and left blank otherwise. Language and currency alone never set it.
Five new discovery sources look for Japanese and Korean AI agents and MCP servers: a Korean MCP index, Japanese and Korean curated lists, Korean makers' own launch posts, Japanese developer articles, and new GitHub projects described in Japanese or Korean.
Product cards for sites in other languages are now written in English. The product keeps its own name in its original script, with an English or romanised form kept for search.

25 Oct 2026

Products people look for on the website and cannot find are now researched for listing

When a visitor searches for a product by name and Agentery does not list it, opens a product page that does not exist, or asks the price verifier for an unknown listing, the name is sent to the same automated research step that MCP lookups already use. If it is a real AI agent, MCP server or agent tool with an official site, it is screened and listed in the usual way; nothing is invented.
Only people using a web browser count. Crawlers, scripts, our own monitoring, test traffic and repeated requests from one address are left out, and task phrases such as "meeting notes" are not treated as product names.
These lookups are used to find missing listings only. They are not counted as demand anywhere.

24 Oct 2026

AEPI v3 coverage update (methodology 1.4.1): more products compared, more sectors ranked

Prices read from another part of the same night's page capture (the full page, a visible price string, or the site's other captured page) are now checked against that text, not only against the extracted pricing section. The evidence standard is unchanged and nothing from another day is used.
About 9% more purchases are compared on an average day (8,403 on 24 October, up from 7,870). A price change found only in that wider text is held for review, not counted, so the update adds coverage rather than new price moves.
Sectors now count each product once, whichever of its listings carried the price that day. 125 of 142 sectors now have enough comparable products to be ranked, up from 86.
The headline reads 100.32 on 24 October (100.37 under methodology 1.4). Four earlier price moves are no longer counted because the comparison now needs review, and one is added. The previous v3 data stays available as the rollback.

24 Oct 2026

AEPI v3: the price index rebuilt from the page archive

The Agent Economy Price Index now runs on AEPI v3. A price change counts only when the same purchase (same plan, billing commitment, currency, included allowance and terms) shows a different payable price and the new price is seen again on the next capture.
The whole history from 30 June 2026 was recalculated from the page captures and price extractions already stored each night. It is not a fresh re-reading of every archived page. Comparisons the evidence cannot settle are left out of that day's index, never counted as unchanged.
Every sub-index, buyer tier, billing lens, protocol slice and sector now uses the same price events and weights as the headline. The headline reads 100.37 on 24 October (the previous version read 100.38); some sub-indices and sectors moved more.
Accuracy so far is model-reviewed only, with no human audit yet. On an average day about a quarter to a third of the weight of comparable purchases cannot be compared and is excluded: a coverage limit.
The previous version stays published and downloadable: aepi.ai/data/index.csv?engine=v2.

24 Oct 2026

Price answers that agree and explain themselves

find_market, market_report, price_benchmark and research_capability now price the same set of products for the same question: products judged to do a different job are left out everywhere, not only in some tools.
Every price answer starts with one plain sentence and states what it covers: how many nearby products, how many do the job, how many publish prices, the buyer tier, the billing basis and the dates of the price observations.
When a buyer tier or delivery type has no verified prices, the answer says so and how many products have prices overall, and shows the broader figures separately, labelled as context. Platforms and APIs are no longer left out of per-type answers.
Demand signals and market gaps leave out leads that come only from your own searches by default; repeated searches by one caller are not independent demand.
Comparisons with a job now name products that only partly do it (with the reason), and the price summary leads with the products that do the job.
Research answers that remain unknown now list exactly what is unverified, which kinds of page were read, and what to ask the vendor.

23 Oct 2026

Retired plans leave listing pages

Listing pages showed every plan Agentery had ever seen, so plans a vendor had withdrawn stayed on the page and could set the entry price and the price verdict. A plan now leaves the page once it has not been on the vendor's pricing page for three days. About 1,800 listings are affected. Past readings stay in our records.

21 Oct 2026

Sector charts list their listings and price changes

When you open a sector on the index charts page, you now see the listings in that sector with their current published prices.
Listings that changed price in the selected period are marked, with the old and new price. Only verified changes are shown, and only from the last 31 days. How much each listing moved the index is still not shown.

20 Oct 2026

Price index history corrected

We checked every price change the index has counted since July against the pricing pages we saved on both days. Most were not real: the page had not changed, and we had read a different figure from it, such as the yearly instead of the monthly price, a 'was' price instead of the current one, or a price converted from another currency.
A first automatic correction this morning still left some false changes in and took some real ones out. So we then re-read every remaining change, and every removed one, by hand against the saved pages. Of the 1,086 changes the index had counted, 333 remain; 41 real changes removed in the morning are back.
The whole index history and every sub-index are recalculated. Today's AEPI is 100.41 (it was shown as 100.61 this morning). Some sub-indices changed more: APIs, for example, now show a fall over the last 30 days instead of a rise. Past daily price-change lists are corrected, and each daily market post now carries a dated correction note.
New price changes are now checked against the saved pages of both days before they count. The automatic checks catch most reading errors but not all, so changes are also reviewed afterwards and corrected if needed.

19 Oct 2026

Sector charts page, and index charts as a grid on desktop

New Sectors page (menu, and a tab on the index pages): every sector price chart, with 24H/7D/1M/3M/All periods, search, and sorting by largest movement, biggest rise or fall, size or name. Thin-evidence sectors can be included.
On desktop, Index charts and Listing growth now open as a grid of full charts, three per row, starting with the headline. Opening a chart shows it large. Phones are unchanged.
Charts no longer stretch their labels on wide screens, and on phones they no longer redraw while loading.
On the desktop homepage, the headline chart is now one card wide, with price movements beside it.

18 Oct 2026

Demand tools fixed, and full product names

demand_signals was returning nothing for almost every caller because it read only about 12 hours of logs. It now reads the last 90 days of recorded external searches that found no provider or only weak matches, shown as general job descriptions with distinct-caller counts (MCP 1.22.1).
market_gaps no longer counts a request a caller made after we had shown that caller the same job, so our own list cannot inflate a gap.
112 listings that had no product name now show the name from their own site. Others no longer show a cut-off last word.
Listings whose site has stopped loading are labelled 'Site not loading'. After 7 days of failures, and a separate plain request that also fails, they are marked inactive and leave search; they come back automatically if the site returns. Inactive listings are hidden from search, listings and current benchmarks, but their past prices stay in the price indices, so historical index values do not change.
The index pages now separate 'Agent connection' (A2A: agents talking to agents) from 'Payment method' (x402, MPP, UCP, AP2, ACP: agents paying). They were one 'Protocol' group. Display only; no index values changed.
New page: Listing growth (/indices/growth) shows how many listings with a price each index, subindex and sector counts, day by day since 6 July, with a CSV download.
From 19 Oct, per-call prices that match a recorded x402 or MPP payment offer count as pay-as-you-go in the index (46 prices from 41 listings). Past index values are unchanged.

17 Oct 2026

Errors that say how to fix the call, and clearer setup

Every MCP and REST error now includes a code, a fix (the corrected call or next step), whether a retry can help, and a docs link. The plain error message is unchanged (MCP 1.22.0).
The MCP instructions now say which tool to use for which question.
The connect page adds setup for VS Code, Codex CLI and Gemini CLI. It also adds a summary of what we record and a short list of promises.
The homepage shows the measured counts behind the numbers: listings, pricing pages read in the latest scan, and the latest signed records. They come from status.json and carry a timestamp.

16 Oct 2026

Signed records: check our data without trusting us

Every price observation and index release now goes into a chained log signed with an Ed25519 key. The first signature covers the full history recorded so far: 906,052 observations and 316 releases.
Anyone can check a manifest, or a listing's current price, at /records using our public key at /.well-known/jwks.json. Signatures contain no prices, and listing-level price history stays private.

16 Oct 2026

New discovery files and a public status page

/llms.txt explains Agentery to language models. /.well-known/ard.json is an Agentic Resource Discovery catalogue of our MCP server and A2A agent, and robots.txt points to it.
/status.json shows measured values only. What we don't measure is null with a reason, and known gaps are listed plainly.

16 Oct 2026

Missing A2A logos fixed

Some A2A listings showed a blank icon, because Google's favicon service had nothing for their endpoint host. 67 now use the icon named in the vendor's own A2A card, and 246 use their own parent domain's icon. About 560 show initials instead of a blank image. New listings are checked daily.

16 Oct 2026

Every A2A listing now shows its A2A check

Product pages and the MCP profile show what we found when we checked the agent's public A2A card and endpoint: card, endpoint, declared version, whether it accepts messages, authentication, messaging cost and skills.
Prices, MCP endpoints and documentation links the vendor publishes in its card are quoted word for word.

16 Oct 2026

Listings now show what vendors tell us directly

When a vendor answers us over A2A, its statements appear on its product page in its own words, with the date and channel. They are the vendor's word, not observed prices, and are never used in AEPI or benchmarks.
Each statement is re-checked weekly against the vendor's public pages. If they differ, we tell the vendor once.

16 Oct 2026

Research can now check MCP and A2A endpoints live

Asking whether a listing exposes an MCP server now gets a live check (connect and list tools) instead of a web-page guess. A2A agent cards are checked the same way.
get_provider accepts provider_id, and request_research accepts a listing URL.

14 Oct 2026

A 90-second introduction on the homepage

First-time visitors see a short film about Agentery. It starts muted; one tap turns the sound on and starts it again from the beginning.
It opens once per browser. Use the 'Watch the 90-second intro' link under the search shortcuts to watch it again.

14 Oct 2026

AEPI and equities: October research update

We published an update to the August study on whether the AEPI leads equity markets: /papers/aepi-equity-lead-lag-update-2026-10-14.html.
The study's SPY result did not hold on data after publication. The 3 September forecast was wrong, and we say so.
A wider full-sample analysis finds a pattern beyond chance that points at growth and AI stocks rather than the whole market. No single setting survives the correction for searching.
Pre-registered tests run from 14 October on the index as published each morning. Results will be reported at 40, 80 and 120 trading days, whatever they show.
aepi.ai/research now has a live out-of-sample tracker. It scores the October signals each trading day from 14 October, using only releases as published before the US open, so you can compare 1, 3, 5 and 10-day settings across 14 ETFs.
Not investment advice.

14 Oct 2026

New index pages, a dark look and protocol coverage

Agentery now has a dark theme across the whole site.
Index charts (/indices) show the headline, every subindex and every sector chart, with 24-hour, 7-day, 1-month and 3-month changes.
The new Index summary (/indices/summary) lists levels and changes for any of the last seven reading dates, and you can download it.
The homepage ranks the largest sector price moves for the period you choose. You can also browse subindex charts by product type, protocol, billing model or buyer tier.
Listings are tagged with the agent protocols they support, each backed by evidence: MCP, A2A, x402, MPP, UCP, AP2 and ACP. From tonight's update there are new UCP, AP2 and ACP indices, and the x402 and MPP indices cover more listings.
Open-source listings with no paid plan are now labelled “Free, self-hosted”, and sites that say they are free are labelled Free.

13 Oct 2026

Search finds the word you typed

One- and two-word searches now also include listings whose name or description contains your word, after the closest matches. Plurals count too: 'postcards' finds 'postcard'.
Search results now show the sector most of them belong to, with that sector's price trend.

12 Oct 2026

Monthly billing checked, not assumed

Some pricing pages switch between monthly and annual prices and show only one at a time. Where we could not confirm which one we saw, the price is now marked 'billing period unverified'.
When you ask for monthly billing, plans that need an annual commitment or whose billing period is unverified are no longer counted as within budget; they are listed separately so you can check them.

12 Oct 2026

Annual billing prices read correctly

Fixed how annual-billing prices are shown. "$9.99 /user billed yearly" next to a $15.99 monthly plan now reads as $9.99 a month on annual billing (MeetGeek showed $0.83/mo).
Yearly totals such as "$1,188 billed annually" next to "$99/month" now read as $99 a month on annual billing, not $1,188 a month. 22 listings were corrected; the price index was not affected.
New pages: terms of service (/terms) and support (/support). The site footer now links Terms and Privacy.

11 Oct 2026

Clearer MCP tool labels

Every tool on the Agentery MCP server now has a readable title and states whether it only reads data or makes a change, so AI assistants can ask before running a write.
The partner-only workflow ranker is no longer listed to public callers; partners with a key still see and use it.

11 Oct 2026

Daily index changes at a glance

The headline index now shows three decimals and its latest daily change beside the value, with the update date in the chart footer. A small but real move is shown with its sign, never as unchanged.
New “What changed?” table at the top of Indices: every headline index and subindex (and all sectors), with latest level, daily, 7-day and 30-day changes, a Last 7 days view of each day's change, search, sorting and CSV download.
Changes use exact calendar dates from the published release; a missing date shows as unavailable rather than zero. If a day's release is late, the page says so and keeps the last reading with its date.

11 Oct 2026

Save products to your watchlist

Tap ☆ on any product (market lists, search results, product pages) to save it. Saved products and watched sectors appear together on a new My watchlist page, linked from the menu once you have saved something. The watchlist stays in your browser; nothing is sent to Agentery.
Product logos are larger, and light logos that disappeared on white now get a dark tile.

11 Oct 2026

Market pages list their products

Each market page now lists its products with logo, a short description, a link to the product and its website.
The price shown follows the buyer tier selected in the fair-price panel: switch to Team and each product shows its Team price, with products that don't offer that tier listed after. Current published prices only.

10 Oct 2026

Fair market price: currency and usage-rate fixes

Subscription prices listed in other currencies were converted to US dollars twice. A €20 plan counted as about $26 instead of about $22, and yen or rupee prices were understated. They are now converted once. 136 of 506 tier medians changed, typically by about 6%. Groups priced only in USD are unaffected.
Usage medians that were printed per batch (for example $5 per 1,000 calls) were divided by the batch size twice, which made them far too low. They are now correct per single unit.
Optional add-ons are now excluded from every group median, as announced. The fair-price file again records which exchange rates it used. AEPI index values are unaffected.

10 Oct 2026

New site design, Connect and Docs

The homepage, Indices, Find & compare, Research and Connect now use the new design, with black chart panels throughout and a dedicated Subindices page.
Connect has MCP, REST and A2A tabs with copyable setup for common clients, example prompts and a tested request and response for each interface.
Docs were rewritten around setup, available tools, parameters, responses, errors and limits, checked against the live service. A new section explains how the fair market price is calculated: buyer tier, currency, billing basis, sample size and exclusions.
Existing links keep working: provider pages stay at /agents/<handle>, and market pages stay at /market/<slug>. No index values, memberships or prices changed.

10 Oct 2026

AEPI: optional add-ons leave the headline (rule addon-r1)

From the 11 Oct observation day, optional add-ons (a phone number, a custom domain, extra storage sold beside a plan) no longer count in the headline AEPI or in group fair-price medians. Their prices and history are still recorded.
Genuine pay-as-you-go offerings stay in: each flagged line was checked against the vendor's own page to see whether it buys the service itself or supplements another purchase.
The rule is applied to both days of every comparison, so the correction creates no price movement. Published releases are unchanged; the rule version is recorded in each release.

10 Oct 2026

Add-on prices, job relevance and named products: follow-up MCP fixes

An add-on (a phone number, a credit pack, storage, a custom domain) is no longer shown as a product's entry price. 27 listings were affected, among them Chat Thing, whose entry is now its $14 Basic plan, not a $5 phone number.
'Cheaper' alternatives now need a paid base plan on both sides, cheaper on the same basis. Both entry prices are shown.
Search results the fit check says do a different job are moved to a labelled 'rejected matches' list instead of taking a result slot. A product you name is never excluded.
Market reports and price benchmarks now count only listings that do the job before calculating any price. With too little scoped evidence they say so instead of giving a precise figure.
Listings repaired for the May data fault also got rebuilt descriptions, comparison profiles and search embeddings.
Missed-product research no longer rejects products whose sites show a bot check. Airbyte, Fireflies.ai, Dante AI, DevRev, Sider, ManyChat and Voiceflow are now listed.
Every MCP tool call now has a 100-second deadline with step logging, after a cold-restart test showed occasional stalls.

10 Oct 2026

MCP repairs from replaying every real question we have ever received

We re-asked all 1,755 distinct questions outside callers have sent the MCP since June against the live tools, fixed what failed, and re-ran the full set.
Comparisons now open with each requested product marked found, ambiguous or missing. Only the products asked for are compared; nothing is substituted, and suggested corrections never swap in a different product. Optional alternatives are listed separately.
Market gaps for a sector now count only jobs checked to be in that sector, and say so plainly when there is too little scoped evidence or no gap. Suggested alternatives exclude add-ons and connectors for the product (listed separately as works-with).
Fixed a data fault from May that gave 106 listings another product's name, links or description (for example Lindy showed Adobe Firefly). Originals are kept; 67 further listings are queued for review, and new profile updates are now refused when their website is on another domain.
Research questions: our own tests now use a separate daily allowance, a used-up allowance is reported as paused rather than as an unknown answer, and a 'Contact sales' button alone is no longer read as 'cannot buy self-service'.
Comparisons are concise by default (about 14 KB instead of about 100 KB; detail:"full" returns everything), head-to-head questions such as 'Cursor vs GitHub Copilot' return both products, and price_benchmark and get_group_fair_price now say which figure answers a quote check.

9 Oct 2026

Clearer MCP tool descriptions, new worked examples, and listing-data limits

The six longest MCP tool descriptions now start with when to use them and are shorter (about 40%), with their limitations kept. All 24 tools and their inputs are unchanged. create_custom_benchmark now says it is the tool for watching the prices of products you choose.
Connect and Docs now cover checking a quote (get_group_fair_price, then price_benchmark) and asking one question about a product (request_research), with two new worked examples recorded from live runs on 9 Oct. Docs now counts all 24 tools; get_group_fair_price was missing from the list.
Our Smithery listing was rescanned and now shows all 24 tools at version 1.21.0 (it had shown 19 since 21 Sep).
Product pages show the shape of a listing's price chart without dates or prices. On aepi.ai, the full index member list and its download are withheld (counts stay public), repricings show the latest day only, and per-listing contributions are no longer published.

8 Oct 2026

Product-group price league, fair prices per group, and MCP 1.21

New: the product-group price league at /aepi/groups. Listings are grouped by what they do (subject, job and customer), never by price, and each group has its own price index starting at 100. Groups are ranked over the same dates; groups without enough price history stay visible but unranked, and mixed groups are marked 'broad market'.
New: a fair price for every product group, set separately for each buyer tier (individual, pro, team / small business, enterprise): the median entry monthly price in that tier, with its typical range. Shown on each group's page and available through the MCP tool get_group_fair_price, with the payment models the group uses and pay-per-use and one-time medians where at least 3 listings publish them.
get_provider_profile now names the listing's product group, with the group's fair price per tier and its 90-day index change. It no longer returns dated price histories.
Fix: plans sold as one-time purchases (credit packs, lifetime deals) were being counted as monthly subscriptions. They are now treated as one-time prices.
Homepage index charts can now be filtered to A2A agents, and to providers accepting x402 or MPP payments.
MCP 1.20.4: shorter, use-led tool descriptions so an agent can tell sooner when Agentery fits the task.

5 Oct 2026

MCP 1.19.0: research that reads, faster answers, searches that keep the job

request_research now reads the vendor's own pages, its verified code repository (README, file list, Dockerfile, Helm chart and licence files, releases) and, for buying questions, the pricing page's buttons and where they actually lead — then answers the exact question asked with quotes, links and observation dates. Most answers arrive in the same response within about 15 seconds; 'unknown' is still given when the sources don't say.
Comparisons and benchmarks in a conversation now answer at normal speed; optional clarifying questions are prepared in the background and offered on your next call. A changed proposed price or other correction is still recalculated in the same turn.
Searches that mention deployment needs (self-hosted, Kubernetes, Helm, container images, signed artifacts) now find tools that do the job first and show, for each result, which of those needs are supported, contradicted or unknown — instead of returning infrastructure tools that merely mention Kubernetes.
compare_providers flags products that look like a different kind of product for your job before you compare prices, and separates software price, open-source licence (no licence fee under a named licence) and the costs of running it yourself.
Benchmark answers now label their cohort as pricing observations from the market neighbourhood, not a ready-made shortlist.

4 Oct 2026

MCP 1.18.0: conversations that remember the job

Task tools (search_providers, compare_providers, research_capability, price_benchmark, market_report) now return one request_ref and a conversation block: the facts remembered for the job (users needing licences, workload, budget, requirements), at most two optional questions when an answer would change a cost or eligibility, and the exact refine_results call to continue. Pass the same request_ref to another task tool to carry the job from search to comparison to benchmark.
Costs are computed for the stated quantities, plan by plan: seat minimums, included allowances and overage, annual commitments and usage rates. Every requested quantity is checked on its own; a missing allowance stays unknown. Add-ons and AI modules are shown as components, never as a whole-product price.
Sellers get observed competitor offers (with tier, unit and observation date), indicative team costs for per-user offers and a plain statement of where a proposed price sits. No optimal price is inferred.
Tested in a private trial by an independent agent before release. Searches stay fast; the clarification step runs on comparisons, benchmarks and replies.

3 Oct 2026

MCP 1.17.2: price-reading fixes

Replaying 72 genuine MCP requests on a restored backup found prices being misread; fixed: thousands separators (IDR 109.000 is 109,000, not 109), numbers inside words no longer read as prices, '/mo' no longer read as million, plans sold only annually can be the entry price, per-run and per-GB-per-day units kept whole, and 'contact sales' shown where no price is published.
A published usage rate or one-time price now counts as a public price in searches that require one.

3 Oct 2026

How the AEPI is made each night

Gather: from 02:15 UTC we pick about 12,300 listings (every listing with pricing, plus one seventh of the rest, so everything is checked at least weekly), open each vendor's pricing page in a real browser, and save a full copy of every page. Everything after this works from our saved copy, which is also backed up offsite.
Read prices: an AI model reads plan names, prices, currency, billing period and per-seat or usage rates from the saved copy (unchanged pages reuse the previous reading), and the results are loaded into the database.
Check before trusting: we confirm the page text really changed, test whether a price swing is plausible, re-read the page for any price that moved, and ignore flip-flopping pages and promotions. Anything doubtful is held, not published. Then we build one clean price row per listing per day.
Calculate: the index compares the same plan from the same vendor day to day; new or vanished plans don't move it. Each vendor counts equally, then each product, plan and price line within it. Example: if vendor A has five plans that all rise 10% and vendor B has one plan that rises 10%, both move the index by the same amount. The day's change is chained onto the previous level (base 100), for All, Agents, MCPs, four buyer tiers and billing types.
Publish: after an integrity check, one daily release feeds aepi.ai, the agentery.com homepage and the MCP price-index tools, so all three show the same numbers. If the check fails, the previous release keeps showing.

3 Oct 2026

MCP: research answers from the right pages; faster search

Research requests are now answered from the pages that state the answer: pricing and signup pages for free trials, documentation for APIs, tokens, rate limits and connection setup. Each question in a request gets its own answer (yes, no, partly or unknown) with a quoted sentence, the page address and the date it was read.
Search reads requests with the fast rules reader unless the request is genuinely ambiguous (non-English text, comparisons, exclusions, unclear names or systems, budget wording it cannot parse). On replayed real requests the median response time is back to about 0.5 s; a request that needs the model still takes several seconds the first time.

2 Oct 2026

AEPI factual correction (1 and 2 October releases)

What happened: the 1 October AEPI release counted eight price comparisons that, on the saved pricing pages, did not compare the same offering: spacefast Plus +200.6% (flat plan became priced per team with different contents), vybe Team −90% (the plan became a credit-budget selector; $0.01 per credit on both days), ojo Subscription −87.2% (included credits halved), reponse All-in −66.9% (included AI conversations cut from 10,000 to 2,000), youtube_mcp_server Scale +241.4% (a five-times larger credit pack under the same name), globalgov Essentials −63.9% (coverage went from 1 country to 193), and rask Business −66.8% monthly and −66.6% annual (included minutes halved).
Correction: these eight comparisons are excluded from 1 October and kept on record; if a later review establishes a like-for-like comparison, the move is counted then. Genuine changes on the same pages still count (for example rask Creator −35% and Creator Pro −34%, whose included minutes did not change). Scout7 (2 October) remains a confirmed product change. The 1 and 2 October releases were re-issued; the originals are kept.
Headline AEPI: 1 October 100.350 → 100.474; 2 October 100.344 → 100.468. Every affected series on 1 October (the same differences carry into 2 October): APIs Individual 95.315 → 101.270; Platforms Team/SME 97.915 → 99.549; APIs 99.447 → 100.903; Infrastructure Team/SME 98.493 → 99.783; Infrastructure Individual 100.299 → 101.567; Platforms Pro 100.357 → 99.871; Infrastructure 99.733 → 100.168; Infrastructure Pro 100.452 → 100.064; Subscription Team/SME 100.420 → 100.754; All AEPI Team/SME 100.409 → 100.741; Subscription Individual 100.134 → 100.431; All AEPI Individual 100.145 → 100.440; Agents Pro 100.253 → 100.504; Platforms 99.857 → 100.070; One-off 101.260 → 101.053; Agents Team/SME 101.266 → 101.462; Subscription 100.365 → 100.516; Agents 100.349 → 100.497; All AEPI 100.350 → 100.474; MCPs Individual 100.359 → 100.466; Agents Individual 99.656 → 99.631; MCPs 100.457 → 100.462; Subscription Pro 100.458 → 100.455; All AEPI Pro 100.461 → 100.459; Subscription Enterprise 100.531 → 100.532; All AEPI Enterprise 100.501 → 100.502.

2 Oct 2026

AEPI methodology change: review of the individual offering

From the next release, a warning (the plan line-up changed, the pricing page was largely rewritten, or a price moved by half or doubled) never excludes a change by itself. It opens an automated review of that specific offering: what is sold (software, managed service, goods), the pricing basis (per account, seat or team), the billing term, every included allowance and listed benefit, each quoted from the saved pages and checked against the plan's section of the page.
Only a confirmed change of what is sold starts a new comparison baseline. The same offering is compared normally; where billing changed, the like-for-like billing is compared; where only one consumable allowance (credits, conversations, minutes, requests) changed, the price per included unit is compared. Anything not established stays pending: the comparison is set aside and kept, the product's other plans keep counting, and the move is counted once if a later review establishes it.
These checks now apply to every price change, including day-to-day changes of products already in the index. They apply to new releases only; published history is changed only through factual corrections.

2 Oct 2026

AEPI factual correction (2 October release)

What happened: the 2 October AEPI release counted a product change as a price rise. Scout7 replaced a $69/month self-serve credits product with a $1,000–$5,000/month managed service; both had a plan called "Growth", so the change was read as a 36-fold price increase.
Correction: the release was re-issued the same morning with Scout7 treated as a new product (no index move). The original release is kept on record.
Every affected series (2 October only; all other days unchanged): MCPs Team/SME 101.664 → 100.821; Subscription Team/SME 100.856 → 100.420; All Team/SME 100.843 → 100.410; MCPs 100.588 → 100.417; Subscription 100.462 → 100.358; Headline AEPI 100.436 → 100.344.

1 Oct 2026

MCP tools now say whether a server is hosted or self-run, and show the licence

Every MCP listing now carries a deployment model as evidence: hosted (a vendor-published MCP endpoint you connect to), self_hosted (a published install line such as npx, uvx, docker or pip), both, or unknown when we hold no evidence. The verdict comes with the excerpt, the source page, the day it was observed and its age, so you can judge it yourself. Nothing is guessed from the existence of a repository.
search_providers accepts a deployment filter (hosted | self_hosted | both | unknown); each result shows its deployment, the vendor endpoint or install line when held, and a note when the observed price belongs to the vendor platform page rather than the connector itself.
get_provider_profile, compare_providers and suggest_alternatives return deployment, licence (named licences only — MIT, Apache, GPL, BSD, MPL, ISC) and how_to_connect.install / how_to_connect.repository. Prices attached to a different offering than the compared connector are labelled price_applies_to.
First pass over the data we already held: 14,494 MCP listings; hosted 843, self-run 4,706, both 2,708, unknown 6,237; a vendor endpoint for 2,115 listings (7 before), an install line for 7,405, a named licence for 6,121. Discovery now keeps the Glama hosting and licence attributes and the MCP registry's remote endpoints and packages with each candidate. Server version 1.10.0.
MCP, REST and A2A now share one set of answers. A keyless REST layer with an OpenAPI description at /api/v1/openapi.json covers search, provider profiles, comparison, alternatives, price benchmarks and AEPI data; an A2A agent card at /.well-known/agent-card.json (built on the official A2A SDK) exposes the same six functions as skills; the Connect page shows the find → inspect → compare journey on each interface. The transport never changes a result, a price or how a missing price is reported.
New free Watch my tools page (/watch): save a shortlist of providers and get confirmed price-change alerts with previous and current price, billing basis, observation date and an evidence link; small moves on non-USD plans are labelled as likely currency effects, not vendor repricing. Optional push to a ntfy.sh topic, once per event. No account.
A concise weekly AEPI brief now publishes on Mondays at /blog/weekly, with rule-chosen findings, named examples and evidence links; quiet weeks are called quiet and currency-dominated moves are explained as such.
Provider profiles carry three separate evidence checks: price source and verification date, Agent Card fetch, and endpoint check result. Unchecked means unknown; an authentication requirement is reported as reachable, not broken.
Discovery now reads a2aregistry.org in full (489 agents, not the first 50) and adds the Global A2A Registry, AgentPeering, AgentRolodex, Agent Tools, Allagents and Agenstry (CC-BY-4.0, attribution kept) as sources, with provenance on every candidate.

30 Sept 2026

AEPI: price changes that survive a quarantine are now counted

When a listing's prices change in a messy way (several plans moving by different amounts, or a plan added or dropped at the same time), AEPI holds it out of the index for two days. Until now it returned at its new prices as if nothing had happened, so the change was never counted. From the 1 October release, a verified change is chained once when the listing returns, and the event is labelled as a catch-up with the day it was first seen, so it is never read as a change that happened today.
Only plans that are genuinely the same (same plan, billing period, unit and currency, printed on the saved page on both days) are compared. When the plan line-up changed, at least two matching plans and half the price list must be covered; anything less is logged for review rather than counted. Very large single-line moves (over 4x) still need an individual review before they count.
A daily next-day audit now checks every price change against the following days' pages and lists the ones the pipeline did not publish. Eleven listings whose price points had been removed by the checks were restored today. Published index history is unchanged; a repaired series for 22–30 September is under review.

28 Sept 2026

Updates

The price index now reads a compact daily snapshot instead of re-reading the whole price history after every restart. Same numbers (verified identical), far less memory.
Price changes the page-text gate could not prove are now re-checked automatically every day and either published (once the page evidence and three days of consistent readings support them), rejected as parser errors, or listed as unresolved in the continuity log with the day they started. No manual approval step; the gate itself is unchanged.

27 Sept 2026

Updates

Price scans now cover the long tail of listings a seventh at a time every day instead of one large Sunday pass, and the parser processes the roster in blocks. Sunday 27 September's scan was delayed until 09:02 UTC by a file-size failure and an exhausted LLM credit balance; the day was recovered in full from the saved render, the integrity gate passed, and AEPI released 27 September normally.

26 Sept 2026

Updates

412 more listings hosted on GitHub now show their verified official website (repository still linked under Sources), and 185 of them show their own product logo instead of a GitHub avatar. Every change is evidence-backed and reversible by batch; search ranking is unchanged.

25 Sept 2026

Updates

pricing provenance fix for repository-hosted listings
Listings hosted on GitHub whose README publishes a tariff now show that price (87 listings, e.g. per-call rates) instead of "No public price found"; the crawler no longer attaches a hosting platform's own plan sheet (GitHub, Smithery, Hugging Face, npm) to a listing, and where it had, that platform price is no longer shown as the listing's. Index methodology unchanged.

24 Sept 2026

Updates

Verified official websites now shown for a first set of 14 listings (repository still linked under Sources), with 7 product logos replacing GitHub avatars and 1 personal avatar replaced by initials. Every change carries evidence and can be reversed. Search ranking is unchanged: a trial of rebuilding listing descriptions from their websites was tested blind against the current descriptions and did not win, so it was not shipped.
The "Repricings today" card on the agentery.com homepage and the aepi.ai repricings feed showed every price change as 0.0% and printed internal tags such as [price_annual] in the price text for part of 24 September 2026. The percentages were correct in the published release; the feed that both pages read had dropped the field when the new engine went live. Fixed the same day: the feed carries the engine's percentage, internal tags never reach the page, and a change with no comparable percentage shows a dash. Index levels and releases were not affected.

23 Sept 2026

Updates

A failed page read is no longer recorded as a blank price. From 23 September 2026, when a pricing page times out, does not resolve, has a certificate problem or cannot be parsed, the day is recorded as unobserved in a separate read-failure ledger, the listing keeps its last good observation with a visible read-failed stamp, and nothing is written into price history. A page that loads but shows no price is recorded as blank only when the saved page text proves it. A daily integrity gate now checks every scan and states its verdict on the continuity log. The 23 September scan was repaired under this rule: 306 blank rows written by failed reads and 31 unproven blanks were moved out of history into the ledger, 75 of the failed pages were re-read successfully the same afternoon, and both index engines were re-released (v1 release 46 with 51 more pairs, v2 release 47 unchanged).
The price index now shows four billing lenses next to the delivery-type views: Subscription, Hybrid, One-off and Pay-as-you-go. A lens is a horizontal cut, not a category: the same product counts in its delivery type (agent, MCP server, API and so on) and in the billing model its prices follow, so the lens rows and options sit in their own group on aepi.ai and on the agentery.com homepage. Definitions are read from the pricing page itself: a subscription is a compulsory recurring charge; pay-as-you-go is a usage rate that needs no subscription; hybrid is one plan with both a compulsory fee and a metered rate; one-off is a one-time purchase price. One engine rule was narrowed at the same time: when a single rate on a product jumps more than four-fold in a day, only that rate is held for the day, no longer the whole product, so modest price changes on the product's other plans now count. History was rebuilt under these rules from 6 July 2026 and is labelled reconstructed; the previous series remain downloadable.
AEPI now includes validated pay-as-you-go pricing. The All index admits standalone pay-as-you-go rates (usage prices you can buy without a compulsory subscription) alongside subscriptions, weighted provider → product → tariff → charge component. A new PAYG sub-index is published on aepi.ai and agentery.com, in the CSV downloads, the constituents list and the MCP tools (get_index_history / compare_indices scope payg; get_price_index provider_type payg, which previously fell back to all). History has been rebuilt under the new rules from 6 Jul 2026 and is labelled reconstructed; the legacy series stays downloadable with ?engine=v1 and keeps publishing daily.

22 Sept 2026

Updates

A repair pass for agents that use Agentery through MCP, shipped in four small releases today. First, the connection contract: the server's own instructions, the tool descriptions and the Connect and Docs pages now all teach the same path (search for the task, open the strongest profiles, compare the shortlist); price benchmarks can be asked for by describing the capability; every advertised option value is one the server actually accepts; and a search that asks for public prices only returns priced products, saying so plainly when none qualify. Second, answer quality: an empty 'cheaper alternatives' answer now states why (reference price unknown, prices not comparable, or nothing cheaper); a product name is never silently swapped for a similar one; a code-host's own pricing page is no longer treated as a project's price; comparisons say how many of the compared plans carry a numeric price; the demand tools count distinct workloads, exclude internal traffic and no longer suggest that a short list means low market demand; test-labelled outcome reports no longer count as evidence that a provider worked. Third, worked examples recorded from real runs on 22 September now sit on the Connect page with a copy-question button, and a Python client route was added. Fourth, public copy: the homepage states what Agentery does for an agent and shows how many listings were checked for prices in the latest scan next to the total indexed; the provider and market example pages state the basis of every price verdict; the customer-support case study was re-checked against the July record and its unsupported comparisons removed; the blog index no longer lists a daily note twice; the 21 September daily note carries dated corrections about the Context.dev promotional prices and its release number. Nothing here changes the price index.

21 Sept 2026

Updates

Follow-up to this morning's recovery statement. The 21-day hole was re-read a second time with the same model that reads pages every night, so the recovered days are read the way every other day is. Result: 1,258 listing-days across 257 listings now carry a price again; 244 of them are free or usage-priced. TrendSpider shows its $82 Standard plan on every day from 8 September to today. Three safeguards came out of the re-read. First, the reader's answer length was capped too low, which silently cut long plan lists in half and produced 'no price' for pages that plainly show one; the cap is raised and a cut answer is now recorded as a failed read, never as 'no price'. Second, a recovered day that disagrees with the genuine reads on either side of it is held out of the index rather than published: 32 such days across 15 listings are held today, each one a case of the reader missing the entry plan or promoting an add-on. Third, a re-read may never replace a day that already passed the page-text check with one that fails it. Today's continuity log at /log was regenerated after these changes: 35 listings went priced-to-blank in today's scan (37 of the 50 with saved page text genuinely show no price now), and the 'missing rows' count no longer counts listings that are only scanned on Sundays.
What broke: on 15 September our price reader could not run for one scan because the reader's credit had run out. That failure was stored as if it were an answer and replayed every morning for pages whose text had not changed, so 166 listings — TrendSpider among them — showed "no public price" for days while their pages plainly showed prices, and the index lost their daily comparisons. What we fixed: we re-read every affected day from the page text we had saved at the time, using the standard reader and the same page-text checks that guard every publication, and restored 832 listing-days (1–21 September) and 209 live price cards. Original records are kept on every corrected row. The index was re-minted from the corrected history on 21 September (release 39; headline levels moved by at most 0.002 points, pair counts rose by about 50 a day) with a provenance note. What changed so it cannot recur: a failed read is never cached; a failed read never replaces a listing's last good price (the day is counted as unobserved, not unpriced); and the public continuity log at /log now raises an alarm when priced listings go blank.
Grok WARP + AgenteryBot excluded from real-users: our own companion agents are now classified as internal on every usage view, including past rows.
A public continuity log now appears at /log after every daily price scan, stating whether the priced-observation chain stayed intact, which listings went from priced to blank, and whether a failed read wrote a blank day into history.
When a search or market lookup finds nothing, the MCP now says why and what to try next instead of returning a bare empty list, and its inputs state plainly that the task or query is required, with an example.
New monthly research series on the blog: The Pricing Ledger. The September edition looks back at how agents, MCP servers and infrastructure were priced between 16 August and 20 September, what changed among continuing listings versus newly discovered ones, and whether similar products charge the same way.

20 Sept 2026

Updates

Price changes now pass a page-text check before publication: the saved pricing page of the day before and the day of must show the new price for the same plan on the same terms, with no list-value, strikethrough, billing-toggle or promotional copy attached. Unclear cases are held for human review instead of being published.

17 Sept 2026

Promotional page variants no longer read as price changes

Some vendors serve two versions of the same pricing page: the list price, and a promotional version (for example "Get 5 months free") that shows only the discounted figure. When our daily read landed on different versions on different days, the index could record the difference as a repricing. One provider, ChartPixel, showed this as a recurring ±67% swing.
The index now recognises a change as a promotion starting or ending when one day’s page carries promotional wording and the price moved by a typical promotion fraction. Such days are kept as observed but do not move the index. A separate nightly check freezes any plan that keeps alternating between two fixed prices to its list price.
We re-ran the last 61 days for ChartPixel and 13 similar listings against stored page evidence: every artificial swing is removed while genuine price cuts and rises still register once. Affected daily points were re-issued (today’s release 25); the overall index level moved by about 0.01 points and the number of tracked plan pairs is unchanged. The rule is versioned and reversible.
Every lesson of this kind is now recorded in our parsing rule registry the same day and loaded into the reader on every run.

3 Sept 2026

Repricings feed: currency noise removed

The repricings lists on the AEPI page and homepage now show only genuine pricing decisions. Small day-to-day wiggles caused by automatic currency conversion (a price drifting from $2.51 to $2.52 with the exchange rate) are kept out of the list, with a note showing how many were set aside.
The AEPI page now has a category summary table under the tier charts — every product category's pair count and 24h / 7d / 30d / 60d movement in one view. Rows are tappable to switch the charts, and the day-stepper arrows re-date the whole table.
The big headline chart now states how many like-for-like pairs back it, same as the tier charts.

30 Aug 2026

One neighbourhood, one number: price trends unified across the site

A reader-style contradiction — a listing showing +12.5% on the edges map but −6.5% on its own market page — exposed that we had two different "markets around a listing": its calibrated substitute neighbourhood, and its broader local market measured by a median. Both were honest; presenting both as 'the market around X' was not.
Fixed structurally: every per-listing price trend anywhere on Agentery is now the same number — a chained, like-for-like index over that listing's accepted commercial substitutes (its own price excluded), with the driving repricing named (e.g. "driven by Zebracat $19→$39"). The median-of-members momentum figure is retired from display; market medians remain as cost levels, which is what they are good at.
Competitive density is now computed on that same substitute set — distinct competing vendors, ranked as a percentile — so the crowdedness axis and the price axis of every chart finally describe the same neighbourhood.
Listings without a public price of their own now get a neighbourhood trend too (their price was never part of the index). Coverage roughly doubled to ~9,900 computed neighbourhoods.
A permanent regression test now asserts that the edges map, the YC page and each listing's market page render identical trend, pair count, vendor count and driver for the same listing.

29 Aug 2026

Y Combinator coverage page, and neighbourhood membership goes cap-free

New: agentery.com/yc — every YC-backed AI-agent company we track (135, batches 2016–2026), folded into the live market graph: which are listed with public pricing, how their competitive neighbourhoods are repricing over 30 days, and how crowded each neighbourhood is. The full cohort is available as JSON at /api/yc. YC status here is provenance on normal listings, not a separate dataset.
Neighbourhood Price Index membership is now threshold-based: a listing's neighbourhood is ALL of its accepted commercial substitutes above the calibrated similarity boundary, with no arbitrary rank cap (previously the closest 30). About a quarter of listings had genuinely larger neighbourhoods than the cap allowed; their indices now average over more like-for-like pairs, which makes extreme readings rarer and better-evidenced. Pair counts remain visible everywhere.
Fixed a duplicate-detection bug in listing intake: a domain could be wrongly matched against any longer domain containing it as text (daily.co vs awardflightdaily.com). Intake now requires exact registered-domain equality.

29 Aug 2026

Index history cleaned of pricing-page scraping artefacts

A reader-flagged anomaly led us to a class of false price changes: pricing pages with billing toggles or currency switches were sometimes scraped in different states, producing phantom back-and-forth "repricings" (e.g. a plan flipping between its monthly and annual-billing price for weeks).
We removed this class from index history: 641 artefact snapshots across 61 providers were flagged and excluded (the underlying data is preserved and the operation is reversible). Over the last 30 days the headline index moved from +0.16% to +0.14% and the enterprise tier from +0.64% to +0.41% — the enterprise story stands, at a more accurate magnitude. The 25 Aug research article carries a revision note.
The same detector now runs in the nightly quality gate, so this artefact class is caught the day it appears — and genuine price changes are never delayed by the check.

28 Aug 2026

Daily market commentary is live on the blog

Every morning, about an hour after the daily price scan lands, a short market commentary is now published automatically at /blog — what repriced, the day's most interesting move charted against its closest market neighbours, 30-day context, and price rises happening in low-competition areas worth watching.

28 Aug 2026

We stress-tested the market engine — and changed nothing

The engine that decides which AI products compete was put through a systematic evaluation: three embedding models, four text-composition templates, and two independent judge models, all measured against nearly 2,000 labelled product pairs.
Every challenger lost or tied under pre-registered rules, so production stays exactly as it is — now validated at 97.4% discrimination accuracy on labelled ground truth. The engine's strength turns out to be its two-profile design (what job a product does + what it can actually do), which is worth about ten times more than the choice of embedding vendor.
Full write-up on the blog: /blog/inside-the-market-engine. Blog articles now have share buttons.

26 Aug 2026

Pricing coverage self-heal: missed public pricing found and fixed

A buyer pointed out that 11x (Alice, Julian) had public pricing we showed as unavailable — their prices live on per-product pages our scanner couldn't see. All seven 11x plans are now listed.
We then measured how widespread this was: of ~3,600 listings marked 'no public price' on their own domains, about 90 turned out to have real pricing pages. 56 of those now show live prices, and the rest are correctly labelled (contact-sales or genuinely unpriced).
A weekly automated check now finds pricing pages our scanner misses, verifies each one actually belongs to that product before accepting it, and keeps them refreshed daily.

25 Aug 2026

First AEPI research article published

New on the site: “Enterprise AI agent prices are quietly climbing — and nobody else's are” — a 31-day like-for-like study showing enterprise-tier prices rising nine times faster than Pro plans. Read it at /blog/enterprise-agent-prices.

24 Aug 2026

Market pages: compact and mobile-first

Pricing now renders as adaptive tier tiles — one row for one tier, side-by-side for two, a single compact row for three, a 2×2 grid for four. Summaries up front (e.g. “Pro · 3 plans · $20–$100/mo”), full plan detail one tap away.
The comparison answer sits above the fold: tight header, Visit/Pricing buttons on one row, pricing summary, tier selector, then the like-for-like comparison.
Market intelligence condensed to two cards — Market density (percentile + strong substitutes) and Price momentum — replacing the taller stat tiles.
Per-type median cards are half the size, two per row, and empty ones no longer render.
Fixed: searching for a product by its name now always finds it. Our search understands what products do, which is great for questions like “invoice processing” — but it meant typing an exact product name could come up empty even when we had the product listed. Name searches now check names, web addresses and product pages directly, on both the website and the MCP connection.
We replayed every past search that wrongly returned nothing — almost all of them now find the right product — and that check now runs automatically every week so this can't quietly come back.
Fixed one listing that was showing another company's name, and audited the whole catalogue for the same problem. A small number of suspect listings were found and are being reviewed by hand.

23 Aug 2026

Clearer market pages: honest tiers, one median language, visible one-time plans

The two “medians” on market pages are now clearly different things: “Median entry price” (stable, across all priced members) vs “Pro-plan median” etc. (follows the tier selector). Same numbers, unambiguous labels.
The tier selector now defaults to the product's own tier, and tiers where the product has no monthly plan carry a subtle marker — they describe the market, not the product.
A product's own displayed price no longer re-labels itself as you explore tiers — it stays pinned (e.g. “$20/mo · only monthly plan”).
One-time packs, usage pricing and custom quotes now appear in the product list (e.g. “$12 one-time · Individual”) instead of reading as “no plan”. They remain excluded from all monthly medians.
Product rows follow the selected tier: plan price at that tier and position vs that tier's median.
One comparison baseline everywhere: the top of a market page now compares a product against its own type's median at the selected tier — the exact same number shown on the benchmark card below. No more two different 'medians' giving opposite verdicts.
Market pages are now full product pages: provider logo, description, Visit website / View pricing buttons, and the product's REAL published plans grouped by tier — including one-time packs, usage pricing and custom quotes, each labelled honestly (shown, but never blended into monthly medians). The plan behind the fixed verdict is marked 'Used for verdict'.
Market comparison is now fully tier-driven: pick a buyer tier and the product's plan, the like-for-like median, the % comparison, spectrum and every product row all follow. Percentages only appear for true like-for-like (same provider type, same tier, monthly-comparable) — missing tiers say 'No [tier] plan', custom quotes and one-time packs say so, and nothing is ever compared across tiers.
Every provider class now gets its own benchmark: platforms compare against platforms, APIs against APIs — not blended into agent/MCP medians. Market pages show a median card per class present, the spectrum colours every class distinctly, and a comparison needs at least two priced peers or it honestly says so.

22 Aug 2026

Faster site, honest market sizes, and a big duplicate cleanup

New on every market page: Market density (unique competing vendors, ranked as a percentile across all Agentery markets) and Price momentum (change in the market's median entry price over the available history window — starting at ~30 days from when our price parser reached full accuracy, widening by a day every day until it stabilises at 90 days). Both use the existing market and pricing methodology; nothing else changed.
Fended off an aggressive AI-training crawler that was fetching over 1,000 pages a minute and slowing the site for everyone. Heavy crawlers are now rate-limited; normal visitors and search engines are unaffected.
Market pages redesigned: one clear overview (what the market is, where the product sits vs the market median), buyer-tier selector, and Agent vs MCP benchmark cards — then the price spectrum and products.
The price spectrum now re-plots per buyer tier: pick Individual/Pro/SME/Enterprise and the dots become the providers with a plan at that tier, at that tier's price — no more tier statistics drawn over unrelated dots.
Market pages now state the TRUE market size — unique vendors and strong substitutes found by capability — instead of a capped list count that made every market look the same size.
Products in a market are grouped by type (Agents, MCP servers, APIs, Platforms) with the most relevant group first, using the same visual language as search.
Search results are now grouped in colour-coded containers — Paid services, Free services, and No public pricing — with slimmer filters.
Removed 57 duplicate listings (the same product listed more than once) and switched on automated duplicate detection: a sensitive weekly scan whose proposed merges must pass a confirmation review before anything is hidden. Nothing is deleted and price data is untouched.

21 Aug 2026

Never a dead end — missing products get researched and listed

Ask for a product we do not have (search or MCP) and it is queued for automated research — real products are verified, screened and listed within ~15 minutes (Apollo.io and Clipto were both added this way today).
New Add listing button in search when a name has no exact match.
MCP tools now say when a name has been queued for research instead of a bare not-found error.
Footer links to the registries carrying our MCP: Smithery, Glama and MCP.so.
Consistent typography and mascot across every page (one font system site-wide).

20 Aug 2026

MCP-first homepage + browse past repricings

Homepage sharpened: Pricing intelligence for agents + MCPs — search, tier comparison, and a Connect Agentery MCP path front and centre.
AEPI section now shows the live listing count, when prices were last updated and when the next scan lands.
Repricings panel gained a day browser — step back through past scan days, with the viewed day marked on the charts.
Docs rebuilt around the current tool set (provider-neutral names, market language) with accurate quickstart examples.

19 Aug 2026

Semantic markets — every listing is its own market

Every listing now has a live market page: its closest competitors by what they actually do, separate Agent vs MCP medians, a price spectrum and buyer-tier breakdowns.
Homepage rebuilt around one search box — type a product or a capability and land in its market.
Provider pages show each product’s position in its own local market.
Fixed category directories retired from navigation in favour of semantic search.

2 Aug 2026

Smarter search + clearer pricing on every niche

Search now treats "free" and "cheap" as price intent, not keywords — a query like "free web-search MCP" returns the market with free / free-tier providers surfaced first, instead of no results.
Niche pages read "free" and "free tier" from the actual pricing classification, so genuinely-free providers are no longer mislabelled as paid or unknown.
The providers table now shows only actionable listings — a published price or a free plan — and each price state is self-explanatory.
The price-distribution chart gained buyer-tier tabs (All / Individual / Pro / Team / Enterprise); pick a tier to compare like-for-like, with the tier median as the line.

1 Aug 2026

Agents and MCP servers, never blended

Agents and MCP servers are priced as two separate markets across the site — niche pages and the homepage example now split the two rather than showing one misleading combined figure.
MCP tools renamed to provider-neutral names (search_providers, get_provider, compare_providers, …). Old agent-named tools keep working as aliases, so nothing breaks.
Rebuilt discovery on free, structured sources — the official MCP registry, PyPI, Glama and the A2A agent registry — after retiring a paid search dependency.

31 Jul 2026

Agent vs MCP price index + Smithery support

The Agentery Economy Price Index (/aepi) and the homepage gained an All / Agents / MCPs scope toggle, each on its own base.
Published a commercial-form pricing classification so the index reflects genuinely commercial listings.
Improved Smithery compatibility: richer server metadata, discovery fixes, and a verified listing link.

30 Jul 2026

Custom benchmarks + broader coverage

Custom Benchmarks (private peer cohorts) — early access: define a cohort and get a reproducible pricing benchmark over MCP.
Began ingesting the A2A agent registry, adding agent-native listings alongside MCP servers.
Honest usage attribution — traffic arriving via Smithery's proxy is labelled as such rather than counted as direct.

23 Jul 2026

Repositioned as a pricing-intelligence engine

Reframed the site and docs around what it's actually for: knowing the going rate and comparing plans and providers for AI agents and MCP servers.
Added a one-click "query this niche over MCP" action on niche pages.
SEO / indexability and first-party analytics so real search demand is measured honestly.

21 Jul 2026

Trustworthy pricing, per buyer tier

Rebuilt the pricing engine to produce corrected per-tier cohorts — currency-normalised, unit-matched, de-duplicated per provider, and honest about small samples (it abstains rather than inventing a range).
Launched the public price index page (/aepi) with day / week / month views and a move-verification bar, so only real repricings show as movements.
New find_niche MCP tool: turn a natural-language task into the canonical market, then its live pricing.

19 Jul 2026

New public site

Redesigned niche pages, provider pages and homepage — cleaner, mobile-first, and built directly on the live data.
A response-quality pass across the MCP tools: clean names, deterministic ordering, and concise summaries by default.

Changelog

Price checks judge each offer on its own evidence

Listings now say when a price may be out of date

Today's price update was late

"MCP endpoint" links on listing pages work again

Very small per-use prices now show in full

Plans marked "coming soon" are no longer shown as current prices

AEPI currency fixes: the vendor's own price is the purchase price (methodology revision 1.6-fx1)

AEPI data: currency attribution, revision dates and data gaps now in every index response

Price changes behind the index: latest release only, product and % change

AEPI methodology 1.6: the index now follows publicly quoted prices, in US dollars

25 more directories feeding new listings, and four sources stopped because their terms forbid it

Agentery MCP: the first answer now says what it is based on and how to sharpen it

More products from China, India, Russia, Brazil, Germany, France, Spain and Latin America, Turkey, Indonesia and Vietnam

Three new ways we find products to list

Sub-indices now show the price changes behind them

Price index: real price changes can now count on the night they appear

Ten products restored that had been removed by mistake

Faster answers from the MCP price tools

Salesforce Agentforce is listed again; hidden listings no longer dead-end a lookup

Same-name products: lookups now show every listing with that name

AEPI methodology 1.5.1: twelve wrongly posted price changes corrected, more products counted

Listings now carry a country, and Japanese and Korean products are discovered directly

Products people look for on the website and cannot find are now researched for listing

AEPI v3 coverage update (methodology 1.4.1): more products compared, more sectors ranked

AEPI v3: the price index rebuilt from the page archive

Price answers that agree and explain themselves

More dependable answers for buying research

Retired plans leave listing pages

Sector charts list their listings and price changes

Price index history corrected

Sector charts page, and index charts as a grid on desktop

Demand tools fixed, and full product names

Errors that say how to fix the call, and clearer setup

Signed records: check our data without trusting us

New discovery files and a public status page

Missing A2A logos fixed

Every A2A listing now shows its A2A check

Listings now show what vendors tell us directly

Research can now check MCP and A2A endpoints live

A 90-second introduction on the homepage

AEPI and equities: October research update

New index pages, a dark look and protocol coverage

Search finds the word you typed

Monthly billing checked, not assumed

Annual billing prices read correctly

Clearer MCP tool labels

Daily index changes at a glance

Save products to your watchlist

Market pages list their products

Fair market price: currency and usage-rate fixes

New site design, Connect and Docs

AEPI: optional add-ons leave the headline (rule addon-r1)

Add-on prices, job relevance and named products: follow-up MCP fixes

MCP repairs from replaying every real question we have ever received

Clearer MCP tool descriptions, new worked examples, and listing-data limits

Product-group price league, fair prices per group, and MCP 1.21

MCP 1.19.0: research that reads, faster answers, searches that keep the job

MCP 1.18.0: conversations that remember the job

MCP 1.17.2: price-reading fixes

How the AEPI is made each night

MCP: research answers from the right pages; faster search

AEPI factual correction (1 and 2 October releases)

AEPI methodology change: review of the individual offering

AEPI factual correction (2 October release)

MCP tools now say whether a server is hosted or self-run, and show the licence

AEPI: price changes that survive a quarantine are now counted

Updates

Updates

Updates

Updates

Updates

Updates

Updates

Updates

Updates

Promotional page variants no longer read as price changes

Repricings feed: currency noise removed

One neighbourhood, one number: price trends unified across the site

Y Combinator coverage page, and neighbourhood membership goes cap-free