# delegate-ai-mcp

MCP servers that let Claude Code delegate light, low-stakes work to local AI CLI backends,
instead of spending its own budget on it. Currently one backend: qwen-code.

## How it works

- `src/qwen-delegate.ts` — core logic: spawns `qwen -p "<prompt>"`, captures stdout, handles
  timeouts and errors. Exported separately from the MCP wiring so it's unit-testable without a
  live MCP connection or a live qwen endpoint.
- `src/qwen-delegate-server.ts` — thin MCP stdio server exposing one tool, `qwen_delegate`.

qwen-code runs natively on Windows here (no WSL), backed by a local OpenAI-compatible model
proxy configured in `~/.qwen/settings.json`. See `research/qwen-mcp-delegation.md` for how that
was confirmed. Replies can take several minutes on a cold local model, so the tool uses a
generous (10 min) timeout rather than trying to enforce "light work" in code — that's a
judgment call left to whoever's calling the tool.

Each backend is its own MCP server process with its own `claude mcp add` registration —
there's no single multi-tool server. Adding a second backend means adding its own
`src/<backend>-delegate.ts` + `src/<backend>-delegate-server.ts` pair and registering it
separately (see Status below — the shared subprocess helper this implies isn't extracted yet).

No build step: everything runs as plain `.ts` source via Node's native TypeScript
type-stripping. No `tsconfig.json`, no `dist/`, no lint/format tooling — deliberate choices,
not oversights (see the wayfinder map linked below).

## Prerequisites

- Node.js 24+ (native TypeScript type-stripping)
- `qwen` on `PATH` and working (`qwen -p "test"` should return a reply)

## Setup

```bash
npm install
npm test
```

## Register with Claude Code

Global (available in every project):

```bash
claude mcp add --scope user qwen-delegate -- node /absolute/path/to/src/qwen-delegate-server.ts
```

Verify it connected:

```bash
claude mcp list
```

## Usage

Once registered, Claude Code can call the `qwen_delegate` tool directly. The routing
convention — *when* to delegate — lives in `~/.claude/CLAUDE.md` under "Delegating light work
to qwen": simple lookups, quick research, boilerplate text, small single-file edits. Claude
decides per-request; there's no enforced dispatcher.

## Testing

```bash
npm test
```

Runs `test/qwen-delegate.test.ts` on Node's built-in test runner (`node --test`) — no test
framework dependency. The subprocess spawn is mocked, so tests run in milliseconds and don't
require a live `qwen` install.

## Status

Project conventions are being worked through a
[wayfinder map](https://git.arthurerlich.de/haylan/delegate-ai-mcp/issues/1) on this repo's
issue tracker (see `docs/agents/issue-tracker.md` for how issues/tickets work here). Decided
so far: no build process, ever; separate MCP server process + registration per backend, with a
shared `src/lib/spawn-cli.ts` helper once a second backend exists (not extracted yet — still
just `qwen-delegate.ts`); no lint/format tooling for now. Still open: a publish-shaped
`package.json`.