Setting up MCP¶
Cohearence speaks the Model Context Protocol (MCP) — an open standard that lets AI clients like Cursor, Claude, Codex, and others connect to external tools and data. Once connected, your assistant can read your Cohearence data, write to it, and run real analysis on it, using everything documented in Ingesting Data and Exposing Analysis Over MCP.
Whichever client you use, the endpoint is the same:
Sign in from your client. Every MCP client we know of now supports signing in directly — you point it at that URL, approve access in the browser once, and it holds the credential for you. There's nothing to copy and no secret sitting in a config file.
A Personal Access Token is the fallback for the one case sign-in can't cover: a runner where nobody can open a browser, even once.
| Signing in | Personal Access Token | |
|---|---|---|
| Best for | Almost everything — Claude, Cursor, Claude Code, Hermes, OpenClaw | Runners with no browser: CI, cloud agents, cron hosts |
| Setup | Approve access once, in your browser | Copy a token from the app, paste it into your client's config |
| Revoking access | Disconnect from the client | Revoke the token from Connect AI at any time |
| Where it lives | Nowhere you manage directly — the client holds a refresh token on your behalf | A config file on your machine (or wherever your agent runs) |
Short version: sign in unless you can't. "Can't" means the agent runs somewhere with no browser and no way to complete a sign-in even once — a CI job, a Cursor Cloud Agent, a cron host — or a client that only accepts a static URL and header.
Continue to:
- Signing In From Your Client — connect Claude, Cursor, Claude Code, Hermes, or OpenClaw without handling a secret.
- Personal Access Token — generate a token for the runners that can't sign in.