Skip to content

MCP setup

MCP (Model Context Protocol) lets an AI assistant such as Claude Code, Cursor or VS Code talk to a tool. If you build with an AI assistant, MCP means it can read RevokeGPT’s docs and, if you allow it, look at and change your sites.

  • You want your assistant to write correct integration code from the real docs: use the docs server.
  • You want it to inspect your sites, analytics and firewall policies, or edit a policy draft: use the management server.

RevokeGPT has two MCP servers: the docs server (no sign-in; search these docs) and the management server (needs a management key, rgmk_...). Set the key once in your shell profile or a secrets manager, never in a committed file. The management key is not your site secret (rgsk_...): a site secret is refused by the management server, and the reverse.

The assistant starts or connects to a server and gets a list of tools. The remote server (https://mcp.revokegpt.com/mcp) is reached over HTTP. The local bridge @revokegpt/mcp (see its reference) runs on your machine over stdio and forwards calls with your key. Tools are read-only except five write tools, and there is no delete, rotate, revoke or publish tool.

  1. Get a management key (rgmk_...) for your organization. Read scopes are on by default; write scopes are opt-in.
  2. Put the key in the environment variable REVOKEGPT_API_KEY.
  3. Add the server to your client (below).
  4. Ask your assistant something small first: “List my RevokeGPT sites.”
Terminal window
# remote
claude mcp add --transport http revokegpt https://mcp.revokegpt.com/mcp
# local (stdio); the variable is expanded when the server starts
claude mcp add --env REVOKEGPT_API_KEY='${REVOKEGPT_API_KEY}' --transport stdio revokegpt -- npx -y @revokegpt/mcp

Remote: Settings, Connectors, Add custom connector, and enter https://mcp.revokegpt.com/mcp. Local: add the stdio server (npx -y @revokegpt/mcp with REVOKEGPT_API_KEY) to claude_desktop_config.json. Claude Desktop does not expand environment variables, so the key goes in that file only.

Use the install link from the install page in the repository (docs/connectors/install.md), or add the same stdio config (command: npx, args: ["-y", "@revokegpt/mcp"], env: { REVOKEGPT_API_KEY: "${env:REVOKEGPT_API_KEY}" }) to your MCP settings.

Open https://docs-mcp.revokegpt.com/install/vscode (or /install/vscode-insiders) and review the entry in the server editor before saving. If ${env:REVOKEGPT_API_KEY} is not expanded, use an inputs prompt in .vscode/mcp.json.

Add a server to ~/.codeium/windsurf/mcp_config.json (serverUrl: "https://mcp.revokegpt.com/mcp", or the stdio config above) and restart Windsurf.

Add the remote URL https://mcp.revokegpt.com/mcp as a connector once it is deployed. See docs/connectors/chatgpt.md in the repository.

You see Meaning What to do
insufficient_scope (HTTP 403) The key lacks the scope this tool needs. Re-consent with the scope, or use a read tool.
unauthenticated Missing or bad management key. Check REVOKEGPT_API_KEY.
rate_limited, busy Too many calls (60 a minute, 4 at once, per connection). Slow down.
Invalid arguments: <path>: <code> The assistant sent input the tool does not accept. Ask it to retry with the fields in the tool’s schema.

The full tool list, scopes and failure codes are in the MCP reference.

  • MCP write tools are listed under the Business plan in the pricing catalog. They also need the write scopes. Leave them off until you need them.
  • Give an assistant the least scope that does the job. Everything it reads (analytics, hostnames) can end up in its context.
  • Never paste a key into a chat, a prompt or a committed file such as .mcp.json; reference the environment variable instead.
  • The server’s text fields from other users or from the web are marked untrusted and sanitized. Still review any change an assistant proposes before you accept it.

The remote server accepts a body of at most 64 KiB, 60 tool calls a minute per connection and 4 in flight at once. These docs are also served to AI tools by the docs MCP server, including llms.txt and llms-full.txt.