MCP server for AI agents

The plumekit CLI ships an MCP server for AI coding agents (Claude Code, Codex, OpenCode and others). It gives them accurate, structured access to PlumeKit: the framework APIs, your project's configuration and the docs, so they write correct code instead of guessing.

What it provides

plumekit mcp speaks the Model Context Protocol over stdio and exposes three tools:

Tool	What it returns
`api_reference(topic)`	An accurate, embedded reference for a core API.
`project_info()`	The current project's `plumekit.toml`: enabled capabilities, targets, build/deploy config.
`search_docs(query)`	Search results from the framework documentation.

api_reference covers these topics: overview, routing, request, response, orm, migrations, forms, views, capabilities, i18n, schedule, helpers, testing, cli, config and portability (the Embedded-Wasm rules for code you write). The agent should call it before writing PlumeKit code.

Run the server from your project root, so project_info finds plumekit.toml, and use the committed ./plumekit wrapper so the agent works at the version your app builds against.

Setup

The MCP config format differs per agent, but each one just runs ./plumekit mcp as a stdio server.

Note: Config keys evolve. If one of these examples is out of date, check the agent's own MCP documentation; the command to run is always ./plumekit mcp.

Claude Code

Add it with the CLI:

Terminal

$claude mcp add plumekit -- ./plumekit mcp

or commit a project-scoped .mcp.json, shared with your team:

JSON

{
  "mcpServers": {
    "plumekit": { "command": "./plumekit", "args": ["mcp"] }
  }
}

Codex

Add an MCP server to ~/.codex/config.toml:

TOML

[mcp_servers.plumekit]
command = "./plumekit"
args = ["mcp"]

OpenCode

Add a local MCP server to opencode.json (project root, or ~/.config/opencode/):

JSON

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "plumekit": {
      "type": "local",
      "command": ["./plumekit", "mcp"],
      "enabled": true
    }
  }
}

Any other MCP client

Any agent that supports MCP stdio servers can run ./plumekit mcp (or plumekit mcp if the CLI is on your PATH). Point it at that command per the agent's MCP config.

Getting the most from it

Tell the agent to use api_reference for PlumeKit APIs before writing code; it keeps generated code accurate. A good project instruction (in AGENTS.md or CLAUDE.md): "This is a PlumeKit app. Use the plumekit MCP api_reference tool for its APIs before writing code."
Run the agent from the project root so project_info finds plumekit.toml. api_reference and search_docs work anywhere: the docs are embedded in the binary, so no framework checkout is needed.
Pair it with the tutorial and CLI reference for the humans on the team.