MCP Tools

The MCP page manages your configured MCP (Model Context Protocol) servers. MCP is a standard protocol that lets the GM call external tools, supporting both local command-line programs and remote HTTP services. It can extend the GM with capabilities such as reading files, querying APIs, and executing searches.

Entry point: “MCP” in the left navigation bar (route: #mcp). On mobile, switch to the “MCP” tab under “Capabilities & Feedback”.

Key Concepts

Transport Types

Each MCP server requires a transport type:

stdio · Local command: Launches a subprocess on the current machine and communicates via standard input/output. The command can be a simple invocation like uvx my-mcp or a full shell command.
http · Remote HTTP: Connects to a running MCP HTTP service at a remote URL (e.g., https://host:port).

Connection Status

The status label at the bottom of each server card has three states:

Connected: The server is enabled and the platform has confirmed it is running.
Not connected: The server is enabled but cannot be reached (process not started or network unreachable).
Disabled: The server is configured but its toggle is off; the GM will not call its tools.

Common Tasks

Adding an MCP Server

Go to the MCP page and click “Add Server” in the top-right.
Enter a “Name” (display name, required).
Select a transport type: “stdio · Local command” or “http · Remote HTTP”.
Enter the “Command / URL” (required): for stdio, the launch command; for http, the service URL.
Optionally, fill in “Environment Variables / Headers” — one entry per line in KEY=VALUE format.
Click “Validate & Enable”. The platform saves the configuration and attempts a connection check.

Enabling or Disabling an MCP Server

Find the server card and toggle the switch in the top-right corner of the card.

On enable, the platform automatically attempts to start the process (stdio) or establish a connection (http).
On disable, the platform stops the server process and the GM will no longer use its tools.

Editing an MCP Server

Click the edit button in the bottom-right of the card to modify the name, transport type, command/URL, environment variables, or Headers. Click “Save” to apply changes.

Deleting an MCP Server

Click the delete button in the bottom-right of the card and confirm. The configuration is permanently removed and cannot be recovered.

Viewing Runtime Logs

Click the logs button in the bottom-right of the card to view the server’s runtime state (process PID and recent stderr output). Click “Export” to download the log as a text file. Admin users see the full stderr; regular users may see only a status summary.

Batch Validation

Click the “Validate” button in the top-right of the page. The platform checks each enabled server’s connection in sequence and shows a toast with the success and failure counts when done.

FAQ

The GM isn’t calling MCP tools? Check all of the following: ① the server toggle is on; ② the card status shows “Connected” rather than “Not connected”; ③ the model in use supports tool calling (some lightweight models do not); ④ for stdio mode, the required command (e.g., uvx) is installed on the local machine and executable.

Status stays “Not connected” after adding the server? For stdio: verify the command runs successfully in a terminal. For http: confirm the target URL is reachable from the server side and is not blocked by a firewall. Click the “Validate” button to retry, or check the logs for specific error messages.

How do I format environment variables? One entry per line in KEY=VALUE format, for example:

API_KEY=sk-xxxxx
BASE_URL=https://api.example.com

For http mode, request headers can be provided in the “Headers” field as JSON, for example {"Authorization":"Bearer xxx"}.

Settings · Models — Configure models with tool-calling capability and API credentials
Modules — Configure built-in platform modules

Was this page helpful?