MCP Server Overview
Connect Claude, Cursor, and other MCP clients to your Coherence workspace.
MCP Server: 2 pages
Coherence ships an official Model Context Protocol server: @coherenceos/mcp-server. It lets MCP clients that run local (stdio) servers — Claude Desktop, Cursor, Cline, Windsurf, VS Code — read from your workspace and ask its AI agent (Nash) to take action. ChatGPT isn't supported yet: its connector needs a hosted MCP server (see the Quickstart).
Source: github.com/getcoherence/mcp-server · MCP Registry: io.getcoherence/mcp
What you can do
Talk to your agent. The headline tool, chat_with_agent, sends a natural-language request to Nash and returns its reply. Nash has its full internal toolset — drafting outreach content (blog articles, social posts, broadcast emails), updating records, posting to social, creating landing pages — all governed by your workspace's approval rules. Posting to social and publishing a landing page require approval before anything goes live. Through MCP, Nash can't send or draft email from a mailbox; it can write the content for you to send. (Agent email, where a workspace has turned it on, works only when you chat with an agent inside Coherence.) From Claude Desktop:
"Ask my Coherence agent to draft a follow-up email to the leads I created this week."
"Have Nash check what's on my calendar today and summarize what I need to prepare for my most important meeting."
Read data directly. For raw lookups where you don't want a full agent round-trip, the server exposes fast-path tools: list_modules, list_records, get_record, create_record, update_record, delete_record, plus the outreach CRUD tools. These hit api.getcoherence.io/v1/* directly and return in under a second.
The full tool list
| Tool | What it does |
|---|---|
chat_with_agent | Send a message to a Coherence agent; agent runs its multi-step tool loop and returns its reply |
list_modules | List CRM modules in the workspace |
list_module_fields | List fields defined on a module |
list_records | List records in a module with search + pagination |
get_record | Get a single record by ID |
create_record | Create a new record |
update_record | Update a record's fields |
delete_record | Soft-delete a record |
list_outreach_items | List blog articles, LinkedIn posts, emails |
get_outreach_item | Get a single outreach item (with MDX content) |
create_outreach_item | Draft a new outreach item |
update_outreach_body | Update the body of an outreach item |
update_outreach_item | Update outreach metadata (title, status) |
For anything not in this list — posting to social, building a website, managing approvals — the right move is chat_with_agent. The agent has those tools and will use the right one.
Security & permissions
Authentication is via an API key (sk_live_…) you create in Settings → API Keys in your workspace. Keys inherit the creating user's permissions and can be optionally scoped:
records:read,records:writecollab:read,collab:writeagents:read,agents:writeworkspace:read
Approval gates (sendEmail, postToSocial, publishSitePage, etc.) are enforced inside the Coherence platform regardless of which scopes a key holds. An MCP user cannot bypass approval flows that exist for human users.
Read more in the Authentication docs →
What's next
- MCP Quickstart — install in your client and run your first prompt
- API Reference — direct REST access without an MCP client
- OpenAPI spec — machine-readable description of every endpoint