Five coding agents average 56% on MCP spec compliance

A compliance tracker at ismcpgoodyet.com puts five coding agents, Codex, Cursor, Claude Code, Grok, and OpenCode, at an average 56% on the Model Context Protocol spec. None of them implement Tasks, the piece of the spec meant to handle long-running work.

That gap sits right where builders need it most. MCP is supposed to be the layer that lets an agent call out to tools and hold state across a session. If Tasks doesn't exist in any client, anything you build that assumes an agent can kick off a job and check back later has nowhere to run.

The rest of the scorecard is uneven in the same direction. The stateless spec revision dated July 28, 2026 gets partial support from only Codex and Claude Code. Cursor has no pagination for tools/list, so a server with a long tool catalog just doesn't page through it. Resources served over skill://, the mechanism for distributing packaged skills through MCP, ship in none of the five clients.

It's not uniformly bad news. In-client OAuth and per-tool filtering are both universal across the five agents, so the basics of connecting a server and scoping what it can touch are solid ground. The holes are specifically in async work and skill distribution, the two areas MCP's own roadmap has been pushing hardest.

If you're wiring an agent to a long-running MCP job right now, you're building on a part of the spec that doesn't exist in the client yet. Check the tracker before you assume a feature works just because it's in the spec document.