MCP
Any Model Context Protocol server becomes a set of tools, indistinguishable from native ones.
Agent("...", mcp=["github", "postgres"])
Agent("...", mcp=["https://tools.example.com/mcp"])
Agent("...", mcp=[{"command": "npx", "args": ["-y", "some-mcp-server"]}])
Registry
These names work out of the box:
| Name | Server | Needs |
|---|---|---|
filesystem |
@modelcontextprotocol/server-filesystem |
|
github |
@modelcontextprotocol/server-github |
GITHUB_PERSONAL_ACCESS_TOKEN |
postgres |
@modelcontextprotocol/server-postgres |
DATABASE_URL |
playwright |
@playwright/mcp |
|
fetch |
mcp-server-fetch |
uvx |
memory |
@modelcontextprotocol/server-memory |
|
slack |
@modelcontextprotocol/server-slack |
SLACK_BOT_TOKEN, SLACK_TEAM_ID |
A missing credential fails at agent construction with the variable's name, not halfway through a run.
Your own servers
nobelagentic mcp add weather "npx -y weather-mcp --units metric"
nobelagentic mcp add internal https://mcp.internal.example.com/mcp
These are saved to ~/.nobelagentic/mcp.json and available in every project by name.
Roadmap
MCP server mode — exposing an agent or tool pack as an MCP server with nobelagentic serve --mcp — is planned for v0.4, so anything you build here plugs into Claude Code, Cursor and other MCP hosts.