MCP integration
Ninja supports the Model Context Protocol. Connected servers expose extra tools the agent can call (with the same confirmation flow as other non-shell tools).
Config locations
| Scope | Path |
|---|---|
| Global (Linux) | ~/.config/ninja/mcp_config.json |
| Global (macOS) | ~/Library/Application Support/ninja/mcp_config.json |
| Global (Windows) | %APPDATA%\ninja\mcp_config.json |
| Project | .ninja/mcp_config.json |
Example
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed"]
},
"my-tools": {
"command": "npx",
"args": ["-y", "my-mcp-server"],
"env": {
"MY_API_TOKEN": "..."
}
}
}
}
| Field | Meaning |
|---|---|
command |
Program to start |
args |
Arguments |
env |
Environment for that process |
Verify
ninja doctor
Doctor reports whether MCP servers connected. A single bad server is usually a warning; Ninja still starts.
Safety tips
- Prefer narrow scopes (limited filesystem roots, least-privilege tokens).
- Store secrets in
envor the process environment — not in chat. - MCP tools still need your approval when they change things (unless you auto-approve).