MCP overview
MCP (Model Context Protocol) lets an assistant call tools against your real data instead of guessing at it. Envpilot runs a remote MCP server, so Claude Code, Codex, or Cursor can look up your projects, read the variables you allow, and fetch a secret file a build needs — through the same auth, scope, rate limits and audit trail as the REST API. Nothing is re-implemented for agents.
Requires the Pro plan (mcp_server) and an API key.
❯terminal
https://www.envpilot.dev/api/mcpTransport is Streamable HTTP; auth is Authorization: Bearer envpk_….
What an agent can do#
| Can | Cannot |
|---|---|
| List projects in the key's scope | Create, edit, or delete anything |
| Read variables and shared accounts | Read outside the key's projects or environments |
Read secret files, with the files resource | Approve its own request |
| Search project names and variable keys | See variable values through search |
File a variable request, with the requests resource | Propose a value for that request |
That last row is the whole trust model: read-only keys, human-approved escalation. An agent states what it needs and why; a human decides and supplies the value. See Agent requests.
Getting connected#
- Mint a scoped key and put it in the client's environment.
- Register the server in Claude Code, Codex, Cursor, or any Streamable HTTP client.
- Confirm with
/mcpthat the tools loaded, not merely that the server is listed.
Limits#
- Pro plan only, re-checked on every call — a downgrade stops the agent immediately.
- Scope is immutable. Projects, environments, resources and surfaces are fixed at creation; widening means a new key.
- Rate limits are shared with REST for that key: 120/min metadata, 30/min value pulls, plus the file buckets. Request filing is 5/hour, burst 2.
- Every value-returning call is audited against the key, including every secret-file fetch.
- Hosted Claude connectors are not supported — they expect OAuth, and Envpilot uses a fixed bearer key. Use Claude Code.