The record
Every number this site publishes, with the query behind it. Click any figure and GitHub recomputes it in front of you — nothing here asks you to take my word for it.
Floors last re-derived 2026-08-10. They are also re-checked automatically, and a claim that stops being true fails the build — see How this page defends itself at the bottom.
Why every number is a floor
You will not find an exact count anywhere on this site. Every figure is written as a floor — “5,000+”, never “~4,800” — because these numbers only grow, and they grow fast enough that precision is a liability. An earlier version of the homepage carried figures that were derived two days before they were written down; throughput had moved by nearly four hundred in between, so the page was inaccurate before anyone read it. A floor does not have that failure mode. It stays true as the number climbs, and the live link carries the exact value so the page never has to.
Directory pipeline throughput
Work on the public plugin directories. The second figure is not a claim about hours spent reading diffs — it is what the automation produced, and it is the strongest evidence that the automation exists.
| Merged pull requests I authored, four public directory repositories | 350+ |
| Pull requests merged through those repositories | 5,000+ |
Broken out by repository
| Repository | Stars | PRs authored | Merged through |
|---|---|---|---|
| anthropics/claude-plugins-official | 33k+ | 210+ | 2,500+ |
| anthropics/knowledge-work-plugins | 23k+ | 50+ | 490+ |
| anthropics/claude-plugins-community | 340+ | 75+ | 2,100+ |
| modelcontextprotocol/mcpb | 2k+ | 10+ | 85+ threads |
| Total | 350+ | 5,000+ |
† The modelcontextprotocol/mcpb figure is open-source maintenance — issue and pull request threads participated in — not pipeline throughput, and it is excluded from the throughput total.
The tool-annotation campaign
Between December 2025 and May 2026 I sent tool-annotation pull requests to MCP servers maintained by other people — adding the metadata that lets a client tell a read-only tool from a destructive one before it calls it. This is the most completely auditable work on this page: every change is a public pull request in somebody else’s repository, and each one had to persuade a maintainer with no reason to accept it.
| Tool-annotation pull requests merged into external repositories | 85+ |
| Distinct external repositories that accepted them | 75+ |
| Combined stars of those repositories | 390,000+ |
The query behind these excludes both accounts I own. That is deliberate and it matters — without those exclusions the same search returns roughly three times as many results, most of them my own repositories, and the number would no longer mean what the sentence above says it means.
Open-source maintenance
Maintainer on the public MCPB repository — the desktop-extension bundle format and its tooling. Contributions span validation, security hardening, and example tooling.
| Issue and pull request threads on the public MCPB repository | 85+ |
What you cannot check here
Most of the engineering runs in private infrastructure and there is no honest way to give you a link to it. The pipelines page describes that work as capabilities rather than artifacts, and you should read it as exactly that: claims backed by the checkable record on this page rather than by evidence of their own. That is the honest arrangement. The public work is not the most important thing I do — it is the part you can fully audit, which is what should make the rest credible.
How this page defends itself
A verification page that goes stale is worse than no verification page at all, so this one is wired to break loudly. Every figure above lives in one module beside the exact query that produces it, and a check re-derives all of them against the GitHub API:
- A live value that falls below a published floor fails the build — the claim has become false.
- A published link that stops resolving fails the build. That sweep runs signed out, because signed in a private repository returns a perfectly healthy 200 and a link broken for everyone else stays invisible.
- A link whose query stops matching the label above it fails the build — a column headed “merged” pointing at a search that does not filter for merged is a quiet overstatement, and it happened here once.
- A floor that has drifted far below reality raises a warning to be revised upward. Floors move up, never down: lowering one to make a red check turn green is falsifying the record, not fixing it.
None of this makes the numbers impressive. It makes them true, which is a different and more useful property, and it is the reason every figure here is worth the click.
Something here look wrong? Tell me. A claim I cannot support is one I would rather delete than defend.