MCP — External Tool Servers
MCP (Model Context Protocol) is an open standard for plugging external tools into the agent — databases, APIs, file systems, dev tools, and your own custom integrations. If a capability exists as an MCP server, AgentX can use it.
How it works
- Open Settings → MCP.
- Add a server: give it a name plus the server command (local
stdioserver) or URL (remote SSE/streaming server). - The server's tools appear in the agent's tool list automatically — no restart needed.
- The agent picks them up like any built-in tool when the task matches.
What you can connect
- Databases — query Postgres, SQLite, etc. through MCP database servers.
- Dev tools — GitHub, GitLab, Docker, Kubernetes MCP servers.
- Files & cloud — Google Drive, S3, local filesystem servers.
- Custom — write your own MCP server (any language) exposing exactly the tools your workflow needs.
Tips
- Start with one server and confirm its tools show up in the tool list before adding more.
- Prefer narrow toolsets — every tool definition costs context on every turn. A server with 50 tools you never use is a context tax.
- Disable, don't delete — if a server misbehaves, toggle it off here; the config stays for later.
- Secrets — MCP servers run with the same secret redaction as the shell: credentials in tool results are masked.
stdio vs remote servers
| Local (stdio) | Remote (URL) | |
|---|---|---|
| Runs | As a process on the phone | On your server / a hosted endpoint |
| Setup | Command + args in settings | Just the URL (+ auth header if needed) |
| Best for | Bundled tools, offline use | Shared team tools, heavy dependencies |