Skip to content

Working with Muse Code

Muse Code is the coding agent used on this repository.

Install and sign in

curl -fsSL https://dev.meta.ai/install.sh -o muse-install.sh   # read it before running it
sh muse-install.sh
muse --version

On Linux the sandbox uses bubblewrap; if muse cannot start it, install the distribution's bubblewrap package. Then, from the repository root:

muse

Choose Sign in with your browser, or paste a Model API key (a separate key from the one in .env keeps the two revocable on their own). Trust the workspace when asked. Muse reads the instructions file at the repository root and nothing above it.

Pay-as-you-go sessions bill per token to the Model API account. Keep the model on the standard IDs and check Billing after the first session.

Working on this repository

  • Run it from the repository root so it sees the whole project.
  • ./dev test is the acceptance check for any change; it runs in Docker, so approve the docker compose command once in the sandbox prompt.
  • Tests run in fixture mode and never spend credits. Ask for a real call only through ./dev check-api once that command exists.
  • .env holds the API key. Keep it out of prompts and out of the files Muse reads.
  • Useful commands: /plan before a larger change, /grilling to pressure-test a design, /taste for a UI polish pass, /goal to pin acceptance checks across turns.

Configuration and models: https://dev.meta.ai/docs/muse-code/