Mcp Server¶

Quickstart¶

Register the server with Claude Code via .mcp.json (the plugin ships this) so the tools appear in your conversation:

{
  "mcpServers": {
    "attune-ai": {
      "command": "uv",
      "args": ["run", "python", "-m", "attune.mcp.server"]
    }
  }
}

Once connected, the built-in tools (code_review, help_lookup, memory_store, …) — plus any registered by installed plugins — are callable from the conversation. To run the server directly for testing:

python -m attune.mcp.server

Tasks¶

Inspect the server's surface from Python¶

Goal: see the registered tools, resources, and prompts without a client.

Steps:

from attune.mcp import create_server

server = create_server()
print(len(server.tools), "tools")
print([r["uri"] for r in server.get_resource_list()])
print([p["name"] for p in server.get_prompt_list()])

Verify: create_server() returns a ready EmpathyMCPServer. server.tools is the merged registry — the 41 built-in tools plus any registered by installed plugins (e.g. attune-redis adds five redis_* tools), so the printed count is ≥ 41. get_resource_list() returns the three attune://… resources; get_prompt_list() returns security-scan / test-gen / cost-report.

Call a tool programmatically¶

Goal: dispatch a tool the way the MCP client would.

Steps:

import asyncio

from attune.mcp import create_server


async def main() -> None:
    server = create_server()
    result = await server.call_tool("auth_status", {})
    print(result)


asyncio.run(main())

Verify: call_tool(name, arguments) is a coroutine — await it. It looks the handler up in the dispatch table and returns the tool's result dict. Rate limiting applies (60 calls / 60 s by default).

Register the server with a client¶

Goal: make the tools available in Claude Code.

Steps: add an mcpServers entry that runs python -m attune.mcp.server (see Quickstart). The plugin's bundled .mcp.json uses uvx --from attune-ai python -m attune.mcp.server; a local checkout uses uv run python -m attune.mcp.server.

Verify: after connecting, the attune tools appear in the client. Server logs land in <tmp>/attune/attune-mcp.log if you need to debug the connection.

Reference¶

The public surface is create_server and EmpathyMCPServer, exported from attune.mcp.

`attune.mcp`¶

Symbol	Purpose
`create_server() -> EmpathyMCPServer`	Build a ready server instance.
`EmpathyMCPServer(...)`	The MCP server (composes `MemoryHandlersMixin` + `WorkflowHandlersMixin`).

`EmpathyMCPServer` — selected members¶

Member	Purpose
`call_tool(tool_name, arguments) -> dict`	Async. Dispatch a tool by name and return its result.
`tools`	The merged tool registry — 41 built-in tools plus any plugin-registered tools.
`resources`	The registered resources.
`get_resource_list() -> list[dict]`	The three `attune://…` resources.
`get_prompt_list() -> list[dict]`	The three prompt templates.
`get_prompt_messages(name, arguments)`	Render a prompt's messages.

Tool-schema groups — `attune.mcp.tool_schemas`¶

Function	Count
`get_workflow_tools()`	21
`get_utility_tools()`	7
`get_help_tools()`	5
`get_memory_tools()`	4
`get_personal_memory_tools()`	4
`get_resources()`	3 resources
`get_prompts()`	3 prompts

Launch¶

Surface	Invocation
Client registration	`.mcp.json` → `python -m attune.mcp.server` (plugin uses `uvx --from attune-ai …`).
Direct	`python -m attune.mcp.server` (stdio).
Python	`create_server()` / `EmpathyMCPServer`.