What the pages showed last month

The month we are looking back on

Between 16 August and 20 September, Agentery’s Sunday full scans grew from 5,673 priced listings to 7,399 (all listings with a website: 18,180 → 22,196), while the daily scans kept reading roughly 8,000 pricing pages a day. That growth is the backdrop for a narrower question the data can actually answer: in how agents, MCP servers (plug-ins that let an AI assistant use a tool), and the infrastructure underneath are priced, what changed over the month, and what stayed the same?

We can answer it two ways. First, by looking at the whole listed market on each Sunday — a snapshot of whoever happened to be in the directory that day. Second, more strictly, by following the 4,186 listings that were priced on both 16 August and 20 September and asking what each of them actually did. The first view mixes composition change with behaviour change. The second isolates behaviour. Both matter; mixing them up is how a quiet month can look like a shift.

Underneath those cuts sit the like-for-like subscription index (AEPI) and the daily archive of published same-plan repricings — whether matched plan prices moved, and how often a same-plan price on a page actually changed. Together with the model-mix tables, they describe a month busy in directory size and plan menus, and quiet in net model switching and matched price levels.

This edition is a look-back: what we observed in the saved pages, then careful interpretation labelled as theory. It is not advice on what anyone should charge.

What we observed
Facts from the pages

Definitions used throughout: subscription = a fixed monthly or yearly fee; usage = a price per unit consumed (token, call, task, credit); one-time = pay once; hybrid = a subscription plus a usage line on the same listing. “Priced” here means at least one plan whose period we could read; free-only and contact-us listings are excluded from every table below, so all denominators match.

The billing-model mix on 20 September

One correction to last week’s numbers first. 719 MCP listings are read from pages hosted on the Smithery marketplace, whose “Hobby free / Pay-as-you-go $10 a month plus $0.10 per 1,000 calls” plan belongs to the host, not to the server. Those pages are shown on their own row so they no longer colour the MCP picture. (Table 1 below uses 718 for the Smithery-hosted row as counted on 20 Sep.)

Table 1 — Billing model by category, listings with a readable plan period, 20 Sep 2026 (count; share of row)

Subscriptions remain the majority model in every category. Usage-only is a real second model for MCP servers and infrastructure (about one in five) and a minority one for agents (one in twelve). The Smithery row shows how a hosting platform can impose a hybrid model on hundreds of listings at once: 40% “hybrid” there is the host’s template, not 289 independent decisions.

The trend across Sundays

Four Sundays give a short trend for agents, MCP servers, and infrastructure. MCP rows in this table include Smithery-hosted pages; the caveat under the numbers matters.

Table 2 — Listings with a readable plan period, by Sunday and category (count; share of that category on that day)

MCP rows include Smithery-hosted pages. Within the month, agents did not move at all (17% with a usage plan on every Sunday; 70–71% subscription-only). Infrastructure was flat too (30–31%). MCP servers went from 30% to 33% with a usage plan — and that is the row where the Smithery template sits: hosted pages were being added through the month, so part of that three-point rise is the template arriving, not servers changing. A fairer reading than “about a point a month” is: flat for agents and infrastructure; a small rise for MCP servers that is partly a hosting artefact.

What existing listings actually did

The cohort of 4,186 listings priced on both endpoint Sundays is the month’s most telling cut. Same listing, same pages — so composition change is out of the picture.

Table 3 — The 4,186 listings priced on both 16 Aug and 20 Sep

Adds and drops of usage pricing are almost equal (91 vs 80), and the two directional switches cancel (12 vs 13). Among vendors we could follow, there was no net movement between models. The month did not produce a one-way migration from subscription to usage, or the reverse. What did move was the menu: nearly one in five listings had a different set of plan names on 20 September than on 16 August — part of which will be our reader seeing the page differently, but much of it is real renaming, retiering, and adding of options.

That distinction — model switch versus menu edit — is easy to blur in a single “pricing changed” headline. Table 3 keeps them separate: renames and retiering show up in the 776; dropping every usage line shows up in the 80 and the 13. This month the second kind nearly balanced itself; the first did not need to.

Who joined the directory this month

New entrants lean more toward usage pricing than the listings that were already there, in every category, by 3 to 7 points. Agents new this month show 19% with any usage plan against 16% for older agents; MCP servers 34% against 29%; infrastructure 36% against 30%. Subscription-only shares among newcomers are correspondingly lower (63%, 53%, 54%) than among older listings (71%, 59%, 62%). Combined with Table 3, that is where the market-level rise comes from: not from vendors switching, but from who is arriving.

Smithery-hosted pages are excluded from Table 4 so the host template does not inflate the “new MCP” usage share. Even without that row, the new-versus-older gap shows in every category. Table 2’s Sunday snapshots therefore embed two stories: a stable incumbent mix, and a slightly more usage-heavy inflow.

Table 4 — Listings first seen 16 Aug–20 Sep vs older listings, model mix on 20 Sep (Smithery-hosted pages excluded)

Price levels over the month

Subscription prices, matched plan for plan, rose about 0.27% over the five weeks (approximate; the index is rebased to 100 at 5 July). From 100.11 on 16 August to 100.38 on 20 September, the path was nearly flat week to week, with the largest step early in September. That is close to flat while the number of matched pairs grew by 28% — from 5,800 to 7,451 — so coverage widened even as the level barely moved.

A growing pair count against a growing directory is the opposite of an index that is quietly losing its subscription base. Whatever else changed on pricing pages this month, the like-for-like subscription comparison set got thicker, not thinner.

Table 5 — Headline AEPI (Agentery’s like-for-like price index, subscription plans only) and the number of day-to-day plan pairs behind it

On the daily side, published same-plan repricings for the fortnight covered by the aepi.ai release archive (beginning 7 Sep) look like this:

Table 6 — Published same-plan repricings per observation day

Two weeks, 58 published repricings on 32 listings: 39 up, 19 down, median size 32%. Daily churn is small: on a typical day 25–40 of roughly 7,500 paired listings show a changed plan price at all, and after the checks a handful survive. (Before 17 September, “held” mostly meant the listing’s market was not yet indexed; from 17 September holds are reserved for suspected artefacts.)

Table 7 — The ten largest published moves of the month, re-read by the new page-text check (live from 21 Sep, applied here retrospectively, dry run)

Five of the ten largest published moves hold up; four are discount layouts read as rises and one cannot be confirmed from the page. They were published before the check existed; corrections are proposed separately and history is kept.

Illustrations of the models we see

As observed on 20 Sep — examples of wording on the pages, not templates to copy:

Do similar products charge the same way?

A natural question after the model-mix tables is whether pricing models travel in families: if a tool charges per call, do the tools that do the same job also charge per call? We can test this because Agentery keeps a semantic map of the directory: every listing has a numeric fingerprint of what it does, and each night we store its 100 closest listings by that fingerprint. For 3,891 listings with a readable pricing type on their own pages (hosting-platform template pages excluded, same-vendor pairs excluded), we looked at the 10 closest neighbours of each and asked how often the neighbour used the same model. The comparison is the share you would expect if neighbours were picked at random.

Table N1 — Share of a listing’s 10 closest neighbours with the same pricing model, 20 Sep 2026

Only one model clearly runs in families: usage pricing. A usage-priced tool’s neighbours are usage-priced more than twice as often as chance, and the result is the same whether we look at 5, 10 or 20 neighbours. Subscription-only tools also sit among subscription-only tools, but only just above chance, because subscription is the default nearly everywhere: an agent’s neighbours are 72% subscription-only whether the agent itself subscribes or not. Hybrid pricing does not cluster at all.

Table N2 — Where the effect is concentrated (MCP servers vs agents)

The tightest family is a specific one: tools that charge a fraction of a cent per call in a stablecoin through the x402 payment protocol (QR-code generators, scrapers, data gateways, signal feeds). 196 of the 807 usage-only listings belong to it, and for those the neighbours are usage-priced 55% of the time. Set that family aside and usage pricing still clusters at twice chance among the remaining tools, so the pattern is broader than one protocol. The opposite cases are just as instructive: listedkit sells document-listing help as credits at $14.99 each while its closest neighbour docjacket sells the same job for $29 a month; genviral sells social content on a $29–$99 monthly ladder while witness-protocol, next to it on the map, charges $0.001 per result.

Possible interpretation
Theories, not findings

The observations above are counts from pages. What follows is interpretation — labelled as theory so it is not mistaken for a finding. The theories try to explain why the market-level picture and the same-listing picture can disagree, and what the quiet index is telling us about the month. None of it is a prescription for what a vendor should charge; this series looks back at what the pages showed, not forward at what anyone ought to put on a page next.

What we will watch next month

• The same-listing switch table (Table 3) on a rolling basis: does the add/drop balance stay even, and do the 12 vs 13 directional switches stay symmetrical?
• Hosted-template share: how many listings’ pricing is set by a platform rather than the vendor, and whether other marketplaces impose templates the way Smithery does.
• Outcome pricing: whether the 38 listings become 50 or 500, and in which categories.
• The page-text check’s daily counts: how many changed plans it passes, rejects, and holds, and how many published moves survive re-reading.
• Per-token price levels: whether the $1.20-per-million median moves as new model versions arrive.
• Menu changes: separate genuine renames and retiering from reader noise by checking the saved page text, and report the real rate.
• Neighbour-vs-pricing lifts (Tables N1/N2): refresh on the last Sunday full scan each month (own pages only, N = 10, Smithery excluded); report lifts, not just shares.