Skip to content
AI

Model Context Protocol

Model Context Protocol (MCP) is how you plug external tools and data sources into Raycast AI. Once a server is connected, its tools become available to AI Chat, Quick AI, and AI Commands alongside everything Raycast already knows about you and your installed extensions.

Install MCP Servers

To add a new MCP server, search for the Install MCP Server command or use the Install New Server action in the Manage MCP Servers command then fill out the form. Give the server a name & icon, select the transport & connection details, optionally a description.

Press Install MCP Server Cmd Enter and Raycast saves the config, starts the connection, and pulls in the tool list automatically. Stdio servers run as soon as they're installed; HTTP servers with OAuth will prompt you to Sign In before any tools become available.

If a stdio command relies on something on your PATH, restart Raycast after editing your shell environment so the new values are picked up.

Configuration Fields

  • Name & Icon: how the server appears in lists, mentions, and the action panel.
  • Transport: Standard Input/Output or HTTP.
  • Command (stdio): the executable to run, e.g. npx or a path to a binary.
  • Arguments (stdio): passed to the command. Accepts a space‑separated string or a JSON array.
  • Environment (stdio): key/value pairs injected into the server's environment, e.g. API tokens.
  • URL (HTTP): the server's MCP endpoint.
  • HTTP Headers (HTTP): key/value pairs sent on every request.
  • OAuth Type (HTTP): Dynamic if the server supports OAuth Dynamic Client Registration, Static if you need to provide your own credentials.
  • Client ID / Client Secret / Scopes (HTTP, Static OAuth only): pre‑registered OAuth credentials and the scopes to request.
  • Description: optional shown in the details pane.

Transports

Raycast supports two ways of talking to an MCP server:

  • Standard Input/Output: a local process that Raycast launches and pipes JSON‑RPC over stdin/stdout. Good for anything you'd run from the terminal: npx, a Python script, a compiled binary.
  • HTTP: a remote endpoint that speaks MCP over HTTP (including the Streamable HTTP variant). Good for hosted services.

Authentication

For HTTP servers that require OAuth, Raycast handles the full flow. Dynamic OAuth uses Dynamic Client Registration with PKCE: click Sign In and Raycast registers itself as a client. Static OAuth uses a Client ID (and optional Secret) you provide, which services like Slack require. Tokens are stored encrypted, per server.

To start fresh, the Logout Server in the Manage MCP Servers action clears saved tokens for an HTTP server without removing it.

Manage MCP Servers

Search Manage MCP Servers in Raycast to see every installed server with its status, tool list, and details in one place. Each server is either Running (connected and serving tools), Stopped (installed but not connected), or in an Error state. The details pane shows the full output from the server or transport.

From this command, you can start a new chat with the selected server or open the Action Panel for more options: Add Server, Refresh, Start, Stop, Restart, Logout, and Uninstall.

Using MCP Servers

Every server gets a @-mention name like AI Extensions. In AI Chat or Quick AI, type @ and enter the server's name to scope a question to it. Type @linear what's assigned to me? to ask the Linear MCP server directly. Raycast also adds an Ask command for each installed server to Root Search (e.g. Ask Linear), prefilling Quick AI with that mention.

Permissions

By default, Raycast asks for approval before running any tool from an MCP server. You can change the default option in Raycast Settings → AI → Permissions, or override it for a single chat from the AI Chat's Chat Settings menu.