Muse Code OpenRouter

Use an OpenRouter API key and any meta/muse* model directly inside Meta Muse Code. After setup, the normal muse command uses meta/muse-spark-1.2 through OpenRouter by default; no Codex configuration is involved.

Why an adapter is needed

Muse Code 0.2.1 exposes a --base-url option but the shipped binary only enables its meta and echo providers. Pointing that Meta provider straight at OpenRouter fails before inference because Muse first requests the proprietary /muse-code/models catalog. Muse also requires a sequence_number on each Responses API stream event, which compatible providers do not always emit.

The local adapter fixes those protocol differences:

• translates OpenRouter's live meta/muse* catalog into Muse's model catalog;
• replaces Muse's endpoint with a loopback-only Responses API adapter;
• supplies the OpenRouter credential without putting it in Muse's Meta auth store;
• preserves the requested meta/muse* model instead of pinning one model;
• clamps Muse's output budget to each advertised model limit;
• adds missing stream sequence numbers; and
• shortens and restores provider-incompatible function names when necessary.

Install

curl -LsSf https://xhluca.github.io/muse-code-openrouter/install.sh | sh

The installer installs Meta's official Muse Code launcher if muse is missing, installs this package for the current user, prompts for the OpenRouter key without echoing it, and then presents a numbered default-model chooser. Afterwards, use Muse Code normally:

muse

List every currently available Meta Muse model and mark the configured default:

muse-openrouter models

Open the chooser again to change the default and refresh Muse's visible picker:

muse-openrouter select

Or select directly:

muse-openrouter select meta/muse-glimmer-30b

Select any listed model for only one run without changing the default:

muse --model meta/muse-glimmer-30b
muse --model meta/muse-spark-1.1
muse --model meta/muse-spark-1.2
muse --model meta/muse-spark-1.2-contributor

Setup and select fetch the live catalog and write every current meta/muse* model as a visible Muse picker entry. Run muse-openrouter select again to pick up newly released models. OpenRouter account privacy settings and provider availability still apply.

Contributor models are clearly marked wherever they are listed or selected. OpenRouter's official disclosure says: Your prompts and outputs may be used to improve Meta's products. They may also be blocked by restrictive OpenRouter data policies.

Manual installation is also supported:

uv tool install muse-code-openrouter
muse-openrouter setup
muse-openrouter doctor --live

For automation, supply the key on standard input so it does not appear in shell history or the process list:

printf '%s\n' "$OPENROUTER_API_KEY" | muse-openrouter setup --key-stdin

What setup changes

• Stores the key at ${XDG_CONFIG_HOME:-~/.config}/muse-code-openrouter/credential, mode 0600.
• Preserves Muse's other settings while updating provider, model, and endpoint_transport in ~/.config/muse/settings.json.
• Writes every current meta/muse* model into Muse's undocumented model_catalog as a visible picker row, including context/output limits and Contributor warnings.
• Saves the original Muse settings under ${XDG_STATE_HOME:-~/.local/state}/muse-code-openrouter/.
• Starts a loopback-only adapter on 127.0.0.1:8817, normally as a systemd user service on Linux.

The adapter never logs API keys, authorization headers, prompts, request bodies, or model responses. Muse session logging remains controlled by Muse itself.

Development

uv run --with pytest pytest
uv run --with ruff ruff check .
uv build

The adapter is dependency-free Python and supports Linux and macOS. The service installer uses systemd when available and otherwise launches a detached process.