Using MCP servers

Blaze can call external tools that speak the Model Context Protocol (MCP) directly from your conversation. For safety, only servers you’ve reviewed and explicitly trusted are ever started.

Configuration file

Server connection details go in ~/.blaze/mcp.json (per environment: production = .blaze, staging = .blaze-stg, development = .blaze-dev). Create the file yourself if it doesn’t exist yet.

{
  "servers": {
    "server-name": {
      "command": "command-to-run",
      "args": ["arg1", "arg2"],
      "env": { "KEY": "value" },
      "env_from": ["ENV_VAR_TO_INHERIT"]
    }
  }
}

• env — environment variables set directly for this server
• env_from — pass through environment variables already set in your terminal. Keep secrets like API keys out of mcp.json itself and pass them via env_from instead

On macOS / Linux, mcp.json must have permissions set to 600 (readable/writable by you only) or it will fail to load.

chmod 600 ~/.blaze/mcp.json

Workflow

1. Check the status of your configured servers

blaze mcp list

2. Review the command, args, and environment variable keys shown, then trust the server if everything looks right

blaze mcp trust <server-name>

3. List the tools a trusted server exposes

blaze mcp tools <server-name>

Once trusted, the AI can call that server’s tools during a TUI or one-shot run.

Changing the config resets trust

Trust is tracked via a fingerprint built from command, args, and the environment variable keys (not values). Changing any of these invalidates the previous trust — the server won’t start again until you run blaze mcp trust once more.

How this interacts with approval modes

In safe mode (the default), MCP tool calls go through the same confirmation prompt as any other tool. Choose “Always allow” on a tool you use often to skip the prompt going forward. See File operation approval for details on the modes.