How we test

Hands-on, client by client, dated. Here's what each label means.

As of 2026-10-14: 37 MCP servers listed, 25 tested hands-on in at least one client, 15 passed every step in all 4 clients; 26 AI agents and 10 AI tools listed. See what changed.

1. Protocol checkWe start the server, do the MCP initialize handshake, list its tools and call one safe tool with known input.
2. InstallWe install it in each client (Claude Code, Cursor, Claude Desktop and ChatGPT (Codex); Codex runs inside the ChatGPT desktop app) with the exact snippet we publish, on a clean profile.
3. Connect & toolsThe client must report it connected within 60 seconds, and its tools must be visible.
4. Agent taskA scripted prompt must make the agent actually call the tool, and the output must pass a concrete check.

Status colors

When we retest

We retest when a tool ships a new major version, when someone reports a problem, or when its link breaks. Our current picks are retested monthly. Every result shows the date it was tested.

We never show “passed” without a date, and we never fill in a result we didn't observe.