Home / Reports
State of MCP connectors, October 2026: what we learned testing 35 remote servers
Published Oct 16, 2026. Numbers on this page update live from our test database. Cite as ConnectorReport, https://connectorreport.com/reports/state-of-mcp-connectors-october-2026.
Key findings
- 55 remote MCP servers tested hands-on. 23 passed every task, 32 passed with caveats, 0 failed outright. Across 338 tasks, 87% passed.
- Real lookups mostly work: 87% of real tasks passed (125 of 144).
- "Nothing found" is the weak spot: only 53% of no-match tests (24 of 45) returned a clean empty result. Most docs and search servers return unrelated results for a nonsense query, so an agent cannot tell a miss from a weak match. Details and fix.
- Input validation is solid: 91% of bad-input tests (40 of 44) got a clean error. The misses were servers that ignored a required field. Details.
- Oversized responses are common: 9 of 55 servers returned 50 KB or more from a single call, up to about 1 MB. Details and fix.
- Slow tools: 3 servers had a task over 10 s (GitMCP for Next.js, Cohere Docs, DeepWiki). Timeout fix.
- Auth: 52 of 55 servers needed no sign-in at all. Of the servers that needed OAuth, GitHub and Slack both failed with "does not support dynamic client registration" in clients that rely on it. Details and fix.
- Protocol versions vary: 43 on 2025-11-25, 6 on 2025-06-18, 4 on 2025-03-26.
- Registry-wide daily handshake: of 1278 public remote endpoints in the official MCP registry, 552 answered without sign-in, 627 answered and asked for sign-in, 56 failed the handshake and 43 were unreachable. That is 92% up and 8% failing.
Top 5 by pass rate and speed
- Hugging Face: 6 of 6 tasks, median 32 ms
- Agentic Terminal Directory: 7 of 7 tasks, median 57 ms
- OpenZeppelin Solidity Contracts: 6 of 6 tasks, median 84 ms
- LLMIntel Model Lifecycle: 5 of 5 tasks, median 170 ms
- Cloudflare Docs: 6 of 6 tasks, median 173 ms
By category
| Category | Tested | Fully passed | Ranking |
|---|---|---|---|
| documentation and docs search | 21 | 5 | Best for documentation and docs search |
| developer tools and code | 13 | 6 | Best for developer tools and code |
| web search and page reading | 5 | 1 | Best for web search and page reading |
| other | 5 | 3 | Best for other |
| public data | 5 | 3 | Best for public data |
| cloud provider docs | 2 | 1 | Best for cloud provider docs |
| team productivity | 2 | 2 | Best for team productivity |
| travel search | 1 | 1 | |
| AI and ML models | 1 | 1 |
Method
Each server ran the same task set with the official MCP TypeScript SDK or MCP Inspector: connect, list tools, two or three real lookups, a no-match query, and bad input. Results are from single runs from a US server on the dates shown on each report. Daily handshakes send an MCP initialize to every public remote endpoint in the official registry. Full data: compare table, incidents, and our MCP server.