The MCP servers real agent setups actually use
Data: 7204 public rigs, 1348 of which (19%) commit at least one MCP server config. Updated 2026-10-09.
Directories list tens of thousands of MCP servers. Very few show up in committed configs. Across the 3346 MCP server entries we parsed, 1130 distinct servers were identifiable (by npm/PyPI package, Docker image or remote host); the rest are local, custom servers.
Top 25
| # | server | rigs | share of MCP rigs | est. tokens |
|---|---|---|---|---|
| 1 | Context7 often with Playwright MCP, Sequential Thinking | 183 | 14% | ~1.2k |
| 2 | Playwright MCP often with Context7, GitHub MCP | 125 | 9% | ~7.5k |
| 3 | GitHub MCP often with Context7, Sequential Thinking | 110 | 8% | ~18.0k |
| 4 | Sequential Thinking often with Context7, GitHub MCP | 70 | 5% | ~900 |
| 5 | Chrome DevTools MCP often with Context7, Playwright MCP | 65 | 5% | ~9.0k |
| 6 | Filesystem often with GitHub MCP, Context7 | 53 | 4% | ~4.2k |
| 7 | Memory (knowledge graph) often with Sequential Thinking, GitHub MCP | 45 | 3% | ~2.6k |
| 8 | Exa often with Context7, Playwright MCP | 39 | 3% | ~1.5k |
| 9 | Notion often with Slack, GitHub MCP | 36 | 3% | ~6.0k |
| 10 | Linear often with Notion, Slack | 35 | 3% | ~5.5k |
| 11 | Postgres often with GitHub MCP, Sequential Thinking | 29 | 2% | ~600 |
| 12 | Figma often with GitHub MCP, Context7 | 28 | 2% | ~3.0k |
| 13 | Serena often with Context7, Sequential Thinking | 28 | 2% | ~8.0k |
| 14 | Atlassian often with Notion, Slack | 25 | 2% | ~10.0k |
| 15 | DeepWiki often with Context7, Playwright MCP | 24 | 2% | ~900 |
| 16 | Slack often with Notion, GitHub MCP | 24 | 2% | ~3.5k |
| 17 | shadcn/ui often with Playwright MCP, Context7 | 23 | 2% | ~2.0k |
| 18 | Fetch often with GitHub MCP, Filesystem | 21 | 2% | ~450 |
| 19 | Sentry often with GitHub MCP, Playwright MCP | 21 | 2% | ~4.5k |
| 20 | Supabase often with GitHub MCP, Playwright MCP | 20 | 1% | ~7.0k |
| 21 | Brave Search often with GitHub MCP, Filesystem | 18 | 1% | ~900 |
| 22 | mcp.grep.app often with Context7, Exa | 15 | 1% | ~2.5k |
| 23 | Microsoft Learn Docs often with GitHub MCP, Context7 | 13 | 1% | ~1.2k |
| 24 | OpenAI Developer Docs often with Context7, GitHub MCP | 13 | 1% | ~1.0k |
| 25 | Puppeteer often with Sequential Thinking, GitHub MCP | 13 | 1% | ~2.2k |
Takeaways
- Docs and browser tools dominate. Look at what's at the top: servers that fetch fresh documentation or drive a real browser — the two things a model can't do from its weights.
- Remote servers are rising. 372 of the identified servers are remote (HTTP/SSE) endpoints with no local install — e.g. GitHub MCP, Exa, Atlassian, DeepWiki.
- Every server has a context price. Tool schemas are sent to the model each turn unless your harness defers them. A big server can cost more than your whole CLAUDE.md — see context tax.