Coherence's MCP server is live
You can now talk to your Coherence workspace from Claude Desktop, Cursor, Cline, Windsurf, and VS Code's GitHub Copilot — AI clients that can run a local Model Context Protocol server. ChatGPT is not supported yet: its MCP connector requires a hosted server, and this release runs locally. The official server is on npm as @coherenceos/mcp-server and listed in the MCP Registry as io.getcoherence/mcp.
The headline isn't that we exposed an API — we already had one. The headline is that you can ask your CRM's AI agent (Nash) to do things from the AI client you already use, and the things it does flow through the same approval rules your team already trusts.
What that looks like
In Claude Desktop, paste seven lines of JSON into the config file, restart, and you can write things like:
"Ask my Coherence agent to draft a follow-up email to the leads I created this week."
"List my open deals over $50k and tell me which ones haven't had activity in 14+ days."
"Have Nash check what's on my calendar today and create a prep task for the most important meeting."
"Add a contact for Jane Doe at Acme Corp, then ask Nash to draft an intro email I can review."
Claude (or Cursor, or Cline) does the orchestration. Nash does the actual work — researching, drafting, creating tasks, updating records — using the same internal toolset Nash has when you talk to it inside Coherence. Drafted emails come back for you to review and send yourself; Nash does not currently send email on its own. The task shows up in your task list. The contact appears in the CRM.
The shape of the integration
Two tiers of tools, by design:
chat_with_agent — the headline. Sends a natural-language message to a Coherence agent and returns the agent's reply. The agent has its full toolset internally: drafting outreach, creating tasks and records, posting to social (behind approval), creating landing pages, reviewing pending approvals. We didn't have to re-expose each of those as a separate MCP tool — Claude calls chat_with_agent, the agent figures out which of its 40+ tools to use, and reports back what it did.
Fast-path REST tools — for direct data access where you don't want a full agent loop. list_modules, list_records, get_record, create_record, update_record, delete_record, plus the outreach content CRUD. These hit api.getcoherence.io/v1/* and return in under a second.
For everything that isn't raw record CRUD — calendar lookups, outreach drafts, social posts, landing page creation, approvals — the right move is chat_with_agent. The agent has those tools and your approval rules still gate the sensitive ones.
How we built it
Most "MCP servers" are thin wrappers over an existing API. Ours is too — about 500 lines of TypeScript over a fresh /v1/* REST surface on our backend. What took the actual work was three decisions that aren't in the source:
1. Per-service auth via a shared library, not a gateway. We considered standing up an Envoy ext_authz filter in front of every service. Then we noticed our prod infrastructure (DigitalOcean App Platform) doesn't run an Envoy gateway — DO's native ingress handles routing. So we extracted the auth code (API key hashing, JWT verification, scope checks) into a @coherence/api-auth package and dropped it into the two services that need it. Linear, Stripe, and Notion all do per-service auth via shared libs. The gateway model is elegant; the library model ships.
2. Approval gates inside the platform, not at the edge. When the agent invokes an approval-gated tool such as postToSocial, the workspace's approval system fires the same way it does for human users. An API key cannot bypass an approval rule that was set up for human users. We could have built a parallel "API-key approval" path; instead we made the existing one universal.
3. Coarse scopes, not per-tool scopes. API keys can be limited to records:read, records:write, collab:read, collab:write, agents:read, agents:write, or workspace:read. Seven scopes covers ~95% of real cases. Per-tool scopes (28+ flags to manage) would have been the "more secure" choice on paper, and the wrong product choice in practice — most users would have left them all on by default.
What's next
This is the stdio transport version — you install via npx and run locally. A hosted HTTP transport is on the roadmap. Until it ships, clients that only connect to hosted MCP servers — including ChatGPT's MCP connector and Claude.ai web connectors — can't use Coherence.
In the meantime, the npx install works in MCP clients that support local stdio servers, including Claude Desktop, Cursor, Cline, Windsurf, and VS Code.
Try it
- Marketing landing: getcoherence.io/mcp
- Quickstart (per-client install): getcoherence.io/docs/mcp/quickstart
- npm:
@coherenceos/mcp-server - Source: github.com/getcoherence/mcp-server
- MCP Registry:
io.getcoherence/mcp
If you've been waiting for "AI tools that actually do the work" — this is what that looks like. Drop the JSON snippet into your config, restart, and ask Nash for something specific. The first time it drafts an email exactly the way you would have written it and all that's left is to review and send it, that's the moment.
