Most directories can tell you what exists. Very few can tell you what happened, because a directory is a photograph and the interesting questions are all about motion.
Sato Hub has been taking that photograph every day since 11 June 2026 — one row per project per day, 100 days of them. That history cannot be reconstructed after the fact. Nobody can go back and find out what a project's README said in June, or whether it has pushed a commit since. You either recorded it or you did not.
Here is what 100 days of recording says about the onchain agent ecosystem, as of 19 September 2026.
1. One in five projects has gone quiet
Of the projects listed on day one that we can measure continuously, 12 of 58 have shown no new public activity in 100 days — no commit, no release, no post.
That is the number a point-in-time directory structurally cannot produce. A snapshot taken today shows a listing; only a series shows a listing that stopped moving. The other 46 are alive and shipping.
Method: projects present in both the 11 June and 19 September snapshots with a recorded last_activity_at on both dates (n=58). "No new activity" means the timestamp is unchanged across 100 days. Projects with no activity signal on either date are excluded rather than assumed dead — which is why the denominator is 58 and not 102.
2. Only 21% of listings have an install anyone has actually run
81 of 383 live listings have a deploy spec whose install was reproduced in an isolated container. The other 302 are documented on the strength of their own README.
This is not an accusation. Most READMEs are fine. But there is a real difference between "the project says it installs like this" and "we ran it and it did", and almost nowhere in this ecosystem is that difference recorded. A model writing setup instructions for you cannot currently tell those two apart.
Method: count of live listings where deploy_spec.deploy_status is verified, or any install variant is. 308 listings have a deploy spec at all; 81 have a reproduced one.
3. 90% of the ecosystem is unverified, and says so
Of 383 live listings, 36 carry an earned verification state. 161 are explicitly self-reported. 186 are unverified.
We publish that ratio rather than hiding it, because the alternative — a checkmark on everything — is how directories become useless. Self-reported is not verified, and the two are never merged on this site.
Method: verification_status across live listings. Only the two evidence-reviewed states count as earned; self-reported and unverified do not.
4. MCP won the interface question
168 of 383 listings implement MCP — more than three times the next standard. x402 has 52, ERC-8004 has 12, A2A has 11.
Eighteen months ago the question of how an agent reaches a tool was open. In this dataset it is not close. The payment layer is where the contest has moved.
Method: the canonical standards field per listing, not tags. A tag means "works with"; the field means "implements". Counting tags instead would inflate the MCP figure and would disagree with our own directory filter.
5. The median project scores 59 out of 100
Across 380 scored listings the median Sato Score is 59: 91 score High, 237 Medium, 52 Low.
The Sato Score measures how open, active and verifiable a project is, from public evidence — never whether it is safe, whether it works, or what it might return. A middling score usually means a real project with a thin public trail: closed source, or quiet, or undocumented, or unreproduced. That describes most of this ecosystem right now.
Method: median of trust_score over live, non-provisional listings. Methodology at satohub.ai/sato-score.
What this does not tell you
These are measurements of public evidence, and public evidence has blind spots. A closed-source project run by a competent team looks quiet here and may not be. A project with a busy repository and no users looks alive. Nothing above is a judgement about whether a tool works, whether a team is good, or whether anything is safe to put funds behind.
The directory grew from 102 listings to 383 over the same 100 days, so the population is not fixed, and the day-one cohort is small enough that the first number should be read as a signal rather than a census.
Every figure is reproducible from the public export. The full catalog is free under CC-BY-4.0 at satohub.ai/api/export/index.json, and the same data is queryable over MCP.