Integrations
Chime integrates with an agent harness through two layers. Only one of them is portable, and that’s by design.
- The portable layer — MCP.
chime mcp serveis a standard Model Context Protocol server. Any MCP-capable host can register it and get the full notify/listen/reply loop with zero Chime-specific code — see MCP for the six tools and the config shape. This is the baseline every harness below meets. - Claude-Code-specific glue. Skill auto-load, a session-start hook that
mints one shared alias, and the Monitor tool (which turns a live
chime receive --watchstream into agent turns) are Claude Code particulars — not part of MCP. They make Claude Code push-notified instead of poll-based; every other harness falls back to the portablechime_wait/chime_receivepattern instead, which works everywhere.
Status
Section titled “Status”Chime is device-verified against real harnesses one at a time, honestly. A harness is Verified only after the full round trip — install, pair, send an approval, receive a reply on the right thread — has actually been run against it. Everything else is Expected: it meets the capability bar on paper (it can run an MCP server or a shell command), but hasn’t been run for real yet.
| Harness | Status | Setup |
|---|---|---|
| Claude Code | Verified — reference integration | Guide |
| Codex CLI | Expected | Guide |
| Gemini CLI | Expected | Guide |
| Cursor | Expected | Guide |
| Aider | Expected | Guide |
OpenClaw and Hermes: verification in progress. Both are planned first-class integrations, but neither has been device-verified yet — no setup guide is published for either until it has been. Once a harness passes the same round-trip test the others did, its guide lands here.
The living, continuously-updated compatibility matrix — including exactly
how each “Expected” row was assessed and when it’s next due for
re-verification — is
VERIFIED-AGENTS.md
in the repo.