muse-openrouter-shim

# muse-openrouter-shim

Run Meta's `muse` coding CLI on any OpenRouter API key. No Meta account, no Meta Model API key, any model you want.

```
muse (unmodified) ──> local shim (port 8399) ──> openrouter.ai/api/v1/responses
```

## Why

Muse Code ships with a single compiled provider that talks to `api.meta.ai` and only accepts a Meta Model API key. The multi-provider code (OpenRouter, OpenAI, Anthropic) exists in the binary but is feature-gated out of the public build.

It turns out the wire protocol is just the OpenAI Responses API plus one proprietary endpoint, `GET /muse-code/models`, that muse needs at startup. So a ~180 line local proxy is enough:

- serves a synthetic model catalog at `/muse-code/models`
- streams `POST /responses` through to OpenRouter untouched (auth passes through, muse never knows)
- maps muse's extended reasoning efforts (`xhigh`, `ultra`) down to `high`
- swallows telemetry so nothing else phones home

The catalog wire format was recovered by pointing `--base-url` at a local listener and iterating on the binary's serde decode errors until the schema parsed. Full field layout is documented in `shim.py`.

## Setup

Requires Python 3 (stdlib only, no pip installs) and an installed `muse` (`curl -fsSL https://dev.meta.ai/install.sh | bash`).

```bash
git clone https://github.com/joymadhu49/muse-openrouter-shim
cd muse-openrouter-shim
cp .env.example .env
# paste your OpenRouter key into .env
```

## Use

```bash
./muse-or                 # interactive TUI
./muse-or exec "fix the failing test"
./muse-or --model x-ai/grok-4.3
```

`muse-or` starts the shim automatically if it is not already running, injects the base URL and model, and hands your OpenRouter key to muse as `META_API_KEY` so the shim can pass it through as the Bearer token. Set a default model with `MUSE_OR_MODEL` in `.env`.

## Model compatibility

Verified working end to end, including tool calls:

| Model | Status |
|---|---|
| meta/muse-spark-1.2 | works (Meta's own model, via OpenRouter billing) |
| openai/gpt-5.1 | works (default) |
| x-ai/grok-4.3 | works |
| google/gemini-2.5-pro | works |
| deepseek/deepseek-v3.2 | works |
| qwen/qwen3-coder | works |
| anthropic/* | blocked: the provider returns a Terms of Service 403 for this scaffold |
| moonshotai/kimi-k2 | fails: 131k context is smaller than muse's startup prompt |

Anything on openrouter.ai/models with a