2026-10-11 16:37 UTC

Favz's maintainers claim their daily census of 166,000+ public GitHub repositories with AI agent configurations โ€” tracking MCP servers, skills, plugins, and hooks โ€” provides a reference corpus for analyzing agent-harness adoption patterns and tooling choices across the open-source ecosystem.

state: watchingheat: mediumuncertainty: mediumnovelscott: mediumagent-harnesses agent-configuration dataset mcp github-censusthirdreplicatorFavz

What is this?

The web results do not directly confirm the specific Favz claim โ€” no result surfaces a 'Favz' project or a 'Show HN: I indexed 166,000 AI agent setups on GitHub' post. The closest signal is an arXiv paper (2603.23802v1, March 2026) analyzing ~177,000 MCP tools from public GitHub repositories, which tracks MCP server creation dates and tool categorization but is an academic snapshot, not a daily census. Other results show general MCP/agent-skills ecosystem activity (AgentConn, SkillsMP claiming ~1.9M skills, GitHub trending discussions) but none attribute a 166k-repo daily census to Favz. The claim remains unverified in the supplied snippets.

Why it matters to Scott

The case claims a daily census of 166k+ GitHub repos with agent configurations (MCP servers, skills, plugins, hooks) โ€” directly engaging Scott's census-vs-snapshot distinction (Retrieval Maturity Ladder), his demand for representative production data in evaluation (Capability Audit, Evaluation-Driven Development), and his work on agent-harness configuration standards and MCP ecosystem tracking. However, the grounding notes the Favz claim is unverified in supplied snippets (no Show HN post or project found), so it remains an unvalidated data-source claim rather than actionable evidence. If verified, it would become a high-value corpus for his agent-config analysis, MCP adoption mapping, and supply-chain security work on auto-installed skills.
ip:concept.retrieval-maturity-ladderip:concept.capability-auditip:concept.evaluation-driven-developmentip:framework.the-index-is-the-data-self-cleaning-wiki-graphip:framework.agent-native-computingip:source.mcp-as-the-tool-belt-standard-giving-ai-agents-hands-and-eyes-ebookip:framework.agent-provenance-stackip:framework.siloosdev:concept.trace-backed-agent-comparisonradar:agentsec-static-config-auditingradar:acs-local-skill-risk-catalogradar:agentstow-coding-agent-config-fanoutradar:ai-agent-security-incidents-datasetradar:agent-memory-add-search-evaluationradar:github-copilot-production-trace-findingsradar:aa-agentperf-local-benchmarkradar:anthropic-project-glasswingradar:casbin-local-agent-policy-gatewayradar:claude-code-skill-load-failures
queries asked of Scott's wikis
  • agent-harness configuration standards (AGENTS.md, mcp.json, hooks/schemas)
  • MCP ecosystem consolidation as USB-C-for-agents pattern
  • dataset quality for agent-tooling adoption analysis โ€” census vs. snapshot
  • open-weight model sovereignty and local inference economics for agent harnesses
  • agent memory / knowledge-graph construction from repo corpora
  • security posture of auto-installed agent skills/plugins (supply chain)

Measured heat

now 0 pts/hpeak 4 pts/hcomments 0/hpeers p14momentum: steady1 platformsage 76h
points/hour across evidence ยท reading as of 2026-10-12 02:59:37.977291+11:00 ยท deterministic, not a model opinion

How the heat travelled

10-08 12:00โญ origin directly observedShow HN: I indexed 166,000 AI agent setups on GitHub
thirdreplicator on hacker news
โ€”
10-08 12:00amplified on hacker news ๐Ÿ‘‘hn.story.50004795
thirdreplicator
peak 2 ยท 0 comments ยท 98% of case engagement
10-08 12:32our radar first saw it ยท +0.5hdiscovery anchor: hn.story.50004795โ€”
pace: p25 vs 1243 stories at the 72h mark (now 76h old) โ€” ahead of aafp-commons-signed-agent-notebook (2.0x), behind agentgate-signed-agent-receipts (0.7x)

Evidence (1) โ€” โญ canonical anchor

sourceobjectauthorscorecomments
๐ŸŸง hn โญShow HN: I indexed 166,000 AI agent setups on GitHub
Retrieved article excerpt

Open article ยท Retrieved 2026-10-08T12:42:06.829154+00:00

# What people really set up in their AI agents

Favz reads the agent config checked in to public GitHub repos and counts the MCP servers, skills and
plugins in it. Not stars, not posts: what is set up. 166,014 repos so far, and it grows each day.

## Rising in the last 30 days

Tools added to the most public repos with agent config, to 2026-10-08. Census numbers, not the sample below.

| Tool | Gain, 30 days | Repos now | Growth |
| --- | --- | --- | --- |
| [release](https://favz.co/tools/skill-release.html) Skill | +299 | 2,295 | +15% |
| [openspec-propose](https://favz.co/tools/skill-openspec-propose.html) Skill +1 in the same pack | +274 | 1,420 | +24% |
| [openspec-sync-specs](https://favz.co/tools/skill-openspec-sync-specs.html) Skill +6 in the same pack | +269 | 1,094 | +33% |
| [openspec-apply-change](https://favz.co/tools/skill-openspec-apply-change.html) Skill +2 in the same pack | +267 | 1,626 | +20% |
| [@playwright/mcp](https://favz.co/tools/mcp-npm-playwright-mcp.html) MCP server (npm package) +1 in the same pack | +266 | 4,594 | +6% |

[All rising tools](https://favz.co/rising/) ยท [RSS](https://favz.co/rising.xml) ยท [Browse every tool](https://favz.co/tools/)



## What well-known repos run

4,340 public repos with 100 stars or more have a page: the tools in their agent config now, and when each came and went. The most starred:

| Repo | Stars | Tools now |
| --- | --- | --- |
| [openclaw/openclaw](https://favz.co/repos/openclaw/openclaw.html) | 391,528 | 1 |
| [microsoft/vscode](https://favz.co/repos/microsoft/vscode.html) | 193,465 | 2 |
| [Significant-Gravitas/AutoGPT](https://favz.co/repos/significant-gravitas/autogpt.html) | 187,681 | 10 |
| [nextlevelbuilder/ui-ux-pro-max-skill](https://favz.co/repos/nextlevelbuilder/ui-ux-pro-max-skill.html) | 132,863 | 13 |
| [electron/electron](https://favz.co/repos/electron/electron.html) | 123,357 | 3 |
| [denoland/deno](https://favz.co/repos/denoland/deno.html) | 108,551 | 6 |
| [pytorch/pytorch](https://favz.co/repos/pytorch/pytorch.html) | 103,714 | 18 |
| [microsoft/playwright](https://favz.co/repos/microsoft/playwright.html) | 96,939 | 0 |

[Browse all repos](https://favz.co/repos/) ยท [Trendsetters: who took tools up early](https://favz.co/trendsetters/)



## Do stars show use?

Only loosely. For 91 MCP servers Favz has both numbers: the stars on the repo, and the GitHub accounts with it in a public repo's agent config. The rank correlation is 0.46, where 1 would mean both rank the tools the same way.

[Stars and use: who is used more than starred, and the reverse](https://favz.co/stars/)



## Let your agent watch this for you

Nobody can keep up with AI tooling by reading. We are building a Favz plugin for Claude Code. It
looks at the tools in your setup and tells you when people with setups like yours start adopting
something new, based on what they installed, not on what is being hyped. It only suggests. It never
installs anything.

The first part works today. Run this in a terminal:

```
npx favz-cli
```

It scans your agent config on your own machine and shows what setups like yours run that you don't.
Point it at one project or a folder of them. It keeps tool names only, and nothing is sent for the
comparison. If you agree to send a short summary, which it shows you first, you get a class and a level,
and a timeline of how your setup grows each time you run it. The source is [on GitHub](https://github.com/thirdreplicator/favz-cli).

The part that checks back on its own is not built yet. Leave your email and we will tell you when it is.

One email when the plugin ships, and one per new report. Nothing else. We store your
email address and the time, and nothing more.

## The census: every repo found so far

Since 2026-09-22 a collector has been looking for every public GitHub repo with AI agent config. This
is not the random sample below. It is everything found so far, and it grows each day. The first full pass over GitHub finished on 2026-10-07. New repos and changes to known ones are still read every day, so the counts keep moving. Built
2026-10-08: 166,014 repos kept out of 211,411 found, and 49,914 tool names that appear in 3 or more repos.

[Browse 3,535 tools](https://favz.co/tools/): each has a page with how it grew and what people run with it. [Browse repos](https://favz.co/repos/) with 100 stars or more: what each one runs and when that changed. [Trendsetters](https://favz.co/trendsetters/): the repos that took tools up before most others did. [Tools by month](https://favz.co/months/): what the most repos first added, month by month.

| Tool | Repos now | Repos ever |
| --- | --- | --- |
| [`mcp:npm:@playwright/mcp`](https://favz.co/tools/mcp-npm-playwright-mcp.html) | 4,594 | 5,080 |
| [`mcp:url:mcp.supabase.com`](https://favz.co/tools/mcp-url-mcp.supabase.com.html) | 4,351 | 4,660 |
| [`mcp:npm:@upstash/context7-mcp`](https://favz.co/tools/mcp-npm-upstash-context7-mcp.html) | 4,335 | 4,846 |
| `agent:code-reviewer` | 3,326 | 3,694 |
| `hook:PreToolUse:bash` | 2,793 | 3,165 |
| [`skill:frontend-design`](https://favz.co/tools/skill-frontend-design.html) | 2,772 | 3,018 |
| `hook:PostToolUse:bash` | 2,573 | 2,955 |
| [`skill:release`](https://favz.co/tools/skill-release.html) | 2,295 | 2,413 |
| `hook:SessionStart:bash` | 2,289 | 2,590 |
| `hook:PreToolUse:node` | 2,206 | 2,432 |
| `hook:PostToolUse:node` | 2,144 | 2,372 |
| [`plugin:frontend-design@claude-plugins-official`](https://favz.co/tools/plugin-frontend-design-claude-plugins-official.html) | 2,092 | 2,278 |
| [`skill:verify`](https://favz.co/tools/skill-verify.html) | 2,008 | 2,104 |
| [`mcp:npm:shadcn`](https://favz.co/tools/mcp-npm-shadcn.html) | 2,002 | 2,159 |
| [`skill:ui-ux-pro-max`](https://favz.co/tools/skill-ui-ux-pro-max.html) | 1,871 | 2,025 |

Adds and removes per day, from commit dates. The last days always look low, because
recently pushed repos are found later. Removes include renames and moves, not only real drops.

| Day | Adds | Removes |
| --- | --- | --- |
| 2026-10-02 | 3,349 | 493 |
| 2026-10-03 | 3,003 | 525 |
| 2026-10-04 | 2,370 | 398 |
| 2026-10-05 | 2,120 | 116 |
| 2026-10-06 | 1,448 | 76 |
| 2026-10-07 | 358 | 96 |
| 2026-10-08 | 7 | 1 |

The same days for the five biggest tools: repos that added (+) or removed (โˆ’) each one.
Every tool's full daily history is in census/events\_by\_day.csv.

| Tool | 10-02 | 10-03 | 10-04 | 10-05 | 10-06 | 10-07 | 10-08 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `mcp:npm:@playwright/mcp` | +5 โˆ’1 | +6 | +4 โˆ’2 | +1 โˆ’2 | +1 โˆ’1 | +1 | +1 |
| `mcp:url:mcp.supabase.com` | +1 โˆ’1 | +1 โˆ’2 | +3 โˆ’2 | +2 โˆ’1 | 0 | +1 | 0 |
| `mcp:npm:@upstash/context7-mcp` | +4 โˆ’4 | +3 โˆ’4 | +3 โˆ’1 | +1 โˆ’1 | +2 โˆ’1 | +1 | 0 |
| `agent:code-reviewer` | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| `hook:PreToolUse:bash` | 0 | 0 | 0 | 0 | 0 | 0 | 0 |

Census files: [census/tools.csv](https://favz.co/census/tools.csv) ยท [census/adoption\_by\_month.csv](https://favz.co/census/adoption_by_month.csv) ยท [census/by\_day.csv](https://favz.co/census/by_day.csv) ยท [census/events\_by\_day.csv](https://favz.co/census/events_by_day.csv) ยท [census/summary.json](https://favz.co/census/summary.json).

## All of public GitHub, counted daily

The sample below is 2,039 repos. This table is wider. Once a day we ask GitHub code search how many
public config files name each tool. These are files, not repos, and the match is on text, so treat
them as rough sizes. Counted 2026-10-08. The change column is since 2026-09-20, the first count, over 18 days. GitHub's search totals wobble by a few percent either way from one day to the next, so a small move means nothing yet. A ranking of what is climbing comes once a few weeks are in.

The last column is the same tool in our random sample. When a tool is large on GitHub and small
in the sample, a project generator probably wrote the config, not a person. Angular's is one.

| MCP servers | Files on GitHub | Change since 2026-09-20 | Repos in sample |
| --- | --- | --- | --- |
| `npm:@angular/cli` | 9,824 | +96 | 11 |
| `npm:@playwright/mcp` | 7,152 | +912 | 138 |
| `npm:@upstash/context7-mcp` | 6,480 | +400 | 113 |
| `url:mcp.supabase.com` | 5,056 | +304 | 63 |
| `npm:@modelcontextprotocol/server-filesystem` | 3,960 | +104 | 34 |
| `npm:shadcn` | 3,240 | +104 | 70 |
| `npm:chrome-devtools-mcp` | 2,864 | +104 | 62 |
| `url:api.githubcopilot.com` | 2,832 | +32 | 51 |
| `npm:@modelcontextprotocol/server-sequential-thinking` | 2,648 | +24 | 32 |
| `npm:@modelcontextprotocol/server-github` | 2,632 | +112 | 31 |
| `npm:@modelcontextprotocol/server-memory` | 1,836 | +106 | 21 |
| `npm:mcp-remote` | 1,792 | โˆ’4 | โ€“ |

| Claude Code plugins | Files on GitHub | Change since 2026-09-20 | Repos in sample |
| --- | --- | --- | --- |
| `frontend-design@claude-plugins-official` | 2,432 | +44 | 35 |
| `superpowers@claude-plugins-official` | 1,692 | +48 | 20 |
| `typescript-lsp@claude-plugins-official` | 1,468 | +64 | 17 |
| `context7@claude-plugins-official` | 1,238 | +42 | 13 |
| `code-review@claude-plugins-official` | 1,206 | โˆ’58 | 11 |
| `code-simplifier@claude-plugins-official` | 1,094 | โˆ’1 | 10 |
| `claude-md-management@claude-plugins-official` | 928 | +34 | 8 |
| `skill-creator@claude-plugins-official` | 854 | โˆ’34 | 12 |

## The first report: what 2,000 real AI agent setups use

Everything from here down is the first report, from a random sample. Its numbers are
separate from the census numbers above.

Draft ยท September 20, 2026 ยท Favz

Directories list what exists. This report looks at what people actually set up. We took a random
sample of 2,039 public GitHub repos that contain AI agent configuration, read the git history of
those config files, and counted tool names. Nothing else was kept.

Most of these repos have no stars at all (1,292 of 2,039; the median is zero). This is what
ordinary builders run, not what famous repos show off.

## Six findings

### 1. Two MCP servers lead, and both serve the same job

| MCP server | Repos that ever set it up | Still there |
| --- | --- | --- |
| Playwright (`@playwright/mcp`) | 151 | 138 |
| Context7 | 138 | 113 |
| shadcn | 72 | 70 |
| Chrome DevTools | 69 | 62 |
| Supabase | 68 | 63 |
| GitHub (hosted) | 60 | 51 |
| sequential-thinking | 43 | 32 |
| filesystem | 41 | 34 |
| GitHub (npm reference server) | 39 | 31 |
| Next.js DevTools | 39 | 35 |

Context7, Supabase and hosted GitHub each appear under two names (a package and a hosted URL). We
merged those. The rest are counted as written.

Four of the top ten (Playwright, shadcn, Chrome DevTools, Next.js DevTools) are for building web
front ends: drive a browser, inspect a page, fetch components. The most common agent setup in public is a web developer's.

0501001502002025-072025-102026-012026-042026-07Playwright 151Context7 138shadcn 72Chrome DevTools 69Supabase 68GitHub (hosted) 60

Running total of sampled repos that had set up each MCP server, 2025-05 to 2026-09. The last month is partial.

### 2. The tools people run together tell the same story

| Pair | Repos with both today |
| --- | --- |
| Playwright + Context7 | 33 |
| Playwright + shadcn | 24 |
| Playwright + Chrome DevTools | 22 |
| Chrome DevTools + shadcn | 20 |
| Next.js DevTools + shadcn | 14 |
| Context7 + shadcn | 13 |
| Context7 + Chrome DevTools | 12 |

One in four Playwright repos also runs Context7. Browser control plus current docs is the closest
thing to a standard stack in this sample.

The map below draws every pairing among the leading MCP servers. Three groups show up. The largest is
web front-end work: Playwright, Context7, Chrome DevTools, shadcn and Next.js DevTools. The second is
the early reference servers, `memory`, `sequential-thinking`, `filesystem` and `github`, which mostly
appear with each other. The third is Microsoft's pair, hosted GitHub and Microsoft Learn.

Even the strongest MCP pairing overlaps by only 20%. People pick MCP servers one at a time. Skills are
the opposite: the design skills in finding 4 overlap by
thirdreplicator20

Interpretation history

Decision trace