Unfragile Research · living census · updated 2026-07-28
The MCP Census
73,954 MCP servers are listed across the four major registries. The registries listing them cannot tell you which ones actually work. So we knocked on every door — automatically, repeatedly, and with a cryptographic receipt for every answer.
The finding
The Model Context Protocol turned every piece of software into a potential tool for AI agents — and cheap intelligence flooded the registries with listings. What nobody built was the checking. A listing is a claim; an agent about to hand a tool its data needs more than a claim.
Our verifier confirmed 14,464 servers alive — a resolvable package on npm or PyPI, or a remote endpoint that answers. It confirmed 27 dead. The remaining 59,463 are repository-only listings that no automated check can currently vouch for — not broken, but unvouched: you would be running them on faith. Only 169 servers in the entire ecosystem clear our highest bar of alive + fresh + adopted + machine-readable contract.
Every confirmed record carries an Ed25519-signed passport — an offline-verifiable receipt of what was checked and when. A chatbot remembers the MCP world as of its training data. This census checked, and it re-checks continuously.
The highest-verified servers
Ranked by verification score: adoption, freshness, liveness, declared contracts, hygiene. Not sponsored; not sponsorable.
| # | Server | Score | Declared tools | Last published | Signed |
|---|---|---|---|---|---|
| 1 | NPM Sentinel MCP | 90 | 19 | 2026-07-24 | ✓ |
| 2 | Qwen3-ASR-0.6B | 87 | 1 | 2026-01-30 | ✓ |
| 3 | Qwen3-ASR-1.7B | 87 | 1 | 2026-01-30 | ✓ |
| 4 | Qwen3-Coder-Next | 87 | 1 | 2026-02-03 | ✓ |
| 5 | Qwen3-Coder-Next-FP8 | 87 | 1 | 2026-02-03 | ✓ |
| 6 | Qwen3-Embedding-0.6B | 87 | 1 | 2026-04-20 | ✓ |
| 7 | Qwen3-Reranker-0.6B | 87 | 1 | 2026-04-16 | ✓ |
| 8 | Qwen3-Reranker-4B | 87 | 1 | 2026-04-16 | ✓ |
| 9 | Qwen3-Reranker-8B | 87 | 1 | 2026-04-16 | ✓ |
| 10 | Qwen3-VL-Embedding-2B | 87 | 1 | 2026-04-16 | ✓ |
| 11 | Qwen3-VL-Embedding-8B | 87 | 1 | 2026-04-16 | ✓ |
| 12 | Qwen3.5-0.8B | 87 | 1 | 2026-03-02 | ✓ |
| 13 | Qwen3.5-0.8B-Base | 87 | 1 | 2026-04-23 | ✓ |
| 14 | Qwen3.5-122B-A10B | 87 | 1 | 2026-04-24 | ✓ |
| 15 | Qwen3.5-122B-A10B-FP8 | 87 | 1 | 2026-04-24 | ✓ |
Methodology — and its limits
Sources: the official MCP registry, Smithery, npm, and Glama, crawled in full and merged by package and repository identity (4,834 servers corroborated by two or more registries). Checks per server: package resolution (npm and PyPI), remote endpoint reachability, repository liveness, declared tool-schema extraction, license presence, and publish freshness. Capabilities are read from declared contracts — actual tool schemas — never inferred by an LLM from marketing copy. Verified records are signed with Ed25519; the public key and verification procedure are published at /trust.
What this census does not yet claim: we do not execute third-party code, so a working clone-and-run server can sit in the unmeasured tier — sandboxed execution testing is the next instrument. Repository-liveness checking is rate-budgeted and proceeds in passes, moving servers out of the unmeasured tier continuously. “Not yet measured” means exactly that — it is not a verdict. When our measurement was wrong, we say so: an earlier internal draft of this census over-claimed, was caught in review, and was corrected before publication.
Use the census
Ask the answer engine — one verified pick, with evidence: GET unfragile.ai/api/answer?q=postgres
Built an MCP server? Claim your record and its signed passport — verify ownership. It’s free.