Connect AI assistants and developer tools to your RudderStack workspace through the hosted Model Context Protocol (MCP) server.
Important
This public repository contains registry metadata, setup examples, and documentation for the hosted RudderStack MCP server. It does not contain the server implementation source code, and its MIT license applies only to this repository's contents.
- Endpoint:
https://mcp.rudderstack.com/mcp - Transport: Streamable HTTP
- Authentication: OAuth 2.0 with browser-based RudderStack login and authorization
- Canonical setup guide: RudderStack MCP documentation
- Tool reference: Read and write capabilities
Native remote-MCP clients can connect directly to:
https://mcp.rudderstack.com/mcp
On the first connection, the client opens a browser so you can sign in to RudderStack and authorize access. No token or client secret belongs in these configuration files.
claude mcp add --transport http --scope user rudderstack https://mcp.rudderstack.com/mcp
Then run /mcp in Claude Code, select rudderstack, choose Authenticate, and sign in to RudderStack in the browser.
In Settings → Connectors, choose Add custom connector, name it RudderStack, and enter https://mcp.rudderstack.com/mcp.
For clients that require a local stdio command, use:
npx -y mcp-remote https://mcp.rudderstack.com/mcp
See examples/.mcp.json for the corresponding JSON configuration.
Until RudderStack is in the ChatGPT app directory, add it as a custom MCP server: open Plugins, select +, choose Add custom MCP server, name it RudderStack, enter https://mcp.rudderstack.com/mcp, choose OAuth, then sign in to RudderStack. If the option is missing, turn on Developer mode under Settings → Security and Login, or ask your workspace admin to allow custom MCP servers. See OpenAI's guide: Connecting custom MCP servers in ChatGPT.
Add to ~/.codex/config.toml:
[mcp_servers.rudderstack] url = "https://mcp.rudderstack.com/mcp"
Then run codex mcp login rudderstack and sign in to RudderStack in the browser.
Create .vscode/mcp.json in your workspace using examples/.vscode/mcp.json, then start the rudderstack server from VS Code's MCP controls.
Copy examples/.cursor/mcp.json to .cursor/mcp.json, then enable rudderstack in Settings → Tools & Integrations → MCP Tools.
In Cline, open MCP Servers → Remote Servers, add a server named rudderstack with URL https://mcp.rudderstack.com/mcp and transport Streamable HTTP, then sign in to RudderStack in the browser when prompted.
Install the extension from this repository:
gemini extensions install https://github.com/rudderlabs/rudderstack-mcp
Then run /mcp auth rudderstack in Gemini CLI, sign in to RudderStack in the browser, and approve access. The extension manifest is gemini-extension.json.
pi mcp add rudderstack --url https://mcp.rudderstack.com/mcp pi mcp login rudderstack
The login opens a browser so you can sign in to RudderStack.
Add a custom MCP server with command npx and arguments -y, mcp-remote, and https://mcp.rudderstack.com/mcp. Clients with native Streamable HTTP support can use the endpoint directly.
RudderStack uses OAuth 2.0 with PKCE and accepts client ID metadata documents alongside dynamic client registration. A compatible client discovers the authorization metadata, opens RudderStack login in your browser, and stores the resulting credentials according to that client's security model.
Tools run with the permissions of the RudderStack user and workspace you authorize. Review tool calls before approving them, especially write operations. Workspace tools require workspace_id (or workspaceId) from user_details on every call; the endpoint is stateless.
Most tools inspect workspace settings, pipelines, schemas, metrics, profiles, audiences, documentation, and service status. Some tools make control-plane changes or trigger syncs, including:
- creating or updating audiences;
- connecting audiences to destinations;
- creating rETL audience activations;
- triggering audience-activation syncs;
- creating or updating transformations; and
- connecting transformations to destinations.
Transformation tests, audience previews, Reverse ETL failed-record reads, and audience-activation planning do not persist control-plane configuration, but they can execute code or submit temporary warehouse work. See docs/tools.md for every tool exposed by the hosted /mcp endpoint, its classification, and conditionally available tools.
- Report suspected vulnerabilities privately as described in
SECURITY.md. - For product bugs, use the bug report form.
- For tool requests, use the tool request form.
- Never post access tokens, workspace IDs, event payloads, customer data, or credentials in a public issue.
The documentation, examples, and metadata in this repository are available under the MIT License. This license does not apply to the hosted server's private implementation.