Connect an MCP Server

MCP (Model Context Protocol) lets an Agent use Tools provided by an external Server. Captain Who currently exposes local stdio MCP Servers to users: the application launches a local program and communicates with it through standard input and output.

Gather the required information

Confirm the following with the MCP Server's trusted publisher:

• the path to the local executable;
• each launch argument and its order;
• the working directory;
• whether the Server will also access the network, files, or other accounts;
• whether the Server's own runtime and dependencies are installed.

Do not copy launch commands from an unknown source, and do not put Tokens, Cookies, or passwords in arguments. The current management UI does not provide configuration for credentials through environment variables, SecretRef, or HTTP headers.

Add a Server

1. Open “Settings → MCP.”
2. In the “External” section, select “Add Server.”
3. Enter a display name.
4. Select the local executable; the transport is fixed to STDIO.
5. Add the arguments one by one in order, then select the working directory.
6. Keep “Allow tools to run automatically” turned off and start with per-call confirmation.
7. Save. A new Server is initially saved in a disabled state.
8. Turn on the enable switch. Review the program, arguments, and working directory in the native confirmation dialog before authorizing launch.

Saving a configuration does not authorize it to run. You must confirm it again if the executable, arguments, working directory, or identified local code entry point changes.

Verify the connection

After enabling the Server, open its details and review:

• connection status;
• launch authorization status;
• whether the Tool Catalog is complete;
• discovered Tools and their security notes;
• the current invocation approval mode.

Only a complete, validated Catalog enters the Agent's available Tool set. Name conflicts, invalid schemas, or oversized content can make some Tools unavailable. Fix the Server first; do not work around the error by weakening approvals.

Approval modes

• Confirm each time: The user reviews and approves every Tool invocation. This is the recommended default.
• Run automatically: Calls that satisfy the current policy may run automatically. Identity and security validation still applies, but unattended side effects may occur.
• Deny invocation: A backend-supported blocking state. If an entry is in this state, its policy must be edited before it can be invoked.

Tool names, descriptions, and readOnly hints supplied by an MCP Server come from a third party. They may aid understanding, but they do not prove that an invocation has no side effects.

Troubleshooting

• Cannot enable: Reopen the details, confirm that the local file still exists, and authorize it again.
• Exits after launch: Check the Server's runtime requirements independently in a terminal, but do not paste sensitive output into a public issue.
• Incomplete Catalog: Check for duplicate Tool names and unusually large descriptions or schemas.
• Invocation times out with an unknown outcome: Do not retry immediately. The external action may already have occurred; verify the target system first.
• Old state remains after an edit: Refresh the MCP page to avoid overwriting a newer configuration.

For more limitations, see Compatibility. For security guidance, see Safe Use.