# MCP server

Source: https://www.fortrabbit.com/features/mcp-server
Created: 2026-08-23
Reviewed: 2026-08-23

> Connect a coding agent once and the platform stops being a browser tab it cannot reach. It looks up which environment runs which PHP version, opens a staging environment, reads the build output of a deployment that failed and sets the variable that fixes it — in the conversation the code was written in.


Connect a coding agent once and the platform stops being a browser tab it cannot reach. It looks up which environment runs which PHP version, opens a staging environment, reads the build output of a deployment that failed and sets the variable that fixes it — in the conversation the code was written in.

## Less switching, fewer copied IDs

The dashboard is a good place to watch what happened. It is a poor place to answer a question mid-task. Copying an app ID between windows to answer one small question breaks the thread every time.

## Authorization that can be taken back

Approval happens in the browser, on a page naming the client that asked. Nothing is pasted into a config file and nothing gets committed by accident. The connection sees exactly what the person who approved it sees, and either side can end it in one step.

## Connect an agent

Setup is one command per client — Claude Code, Codex, Cursor, VS Code — then one approval in the browser. The MCP server is in testing, and the platform around it is in beta.

:BlockLink{title="MCP setup instructions" path="/mcp" property="www"}

Every tool the agent gets, what it may change, and how a connection is revoked:

:BlockLink{title="MCP server documentation" path="/platform/automation/mcp" property="docs"}

## FAQ

### Which agents can connect?

**Anything speaking MCP** — Claude Code, Codex, Cursor and VS Code have a documented setup command each. Any other MCP client works against the same endpoint. A client without OAuth support authenticates with a personal API token instead.

### What can an agent actually change?

**Three things** — Most tools read. Three write: creating an app, creating an environment, and updating environment variables. Restarts, deletions and standalone deployments are not exposed — a REST endpoint existing does not make it available to agents.

### Can an agent see more than the person who connected it?

**Never** — A connection carries the authorization of the person who approved it and sees exactly what that person sees in the dashboard. The access token expires after one hour and the refresh token after 30 days, both bound to that one agent.

### How is a connection revoked?

**From either side** — Clearing the authentication in the agent drops the access token and the refresh token together, so it cannot quietly renew. Disconnecting on the account side ends every connection that agent holds — the route to take when the machine running it is out of reach.

### Is it ready for production work?

**It is in testing** — The MCP server is in testing, and the platform around it is in beta. Reading environment variables returns the values fortrabbit injects, database credentials among them, and whatever an agent reads lands in the context of the connected client. Worth a thought before pointing one at a production environment.

---

- [Agent skills](/features/agent-skills)
- [Agentic engineering](/workflows/agentic-engineering)
- [MCP server docs](/platform/automation/mcp)