muse-execution-backend-python

Meta Muse Sandboxed Execution with E2B (Python)

This example adapts Meta's sandboxed execution cookbook: the Muse Spark agent loop stays in the caller process, while E2B replaces Docker as the disposable execution environment.

The model receives two tools backed by one E2B sandbox:

• run runs a command from the repository workspace.
• write_file writes a file below that workspace.

The Meta API key stays in the caller process. The sandbox is terminated in a finally block after the run.

How to run example

1. Set API keys

Copy the example environment file and set E2B_API_KEY and META_API_KEY. You need access to the E2B dashboard and Meta Model API with Muse Spark.

cd examples/muse-execution-backend-python
cp .env.example .env

E2B_API_KEY=...
META_API_KEY=...

META_BASE_URL defaults to https://api.meta.ai/v1, and META_MODEL defaults to muse-spark-1.2.

2. Install dependencies

uv sync --all-groups

3. Run the agent

The example clones a public HTTPS repository into a fresh sandbox and keeps it alive for the complete agent run.

uv run python -m muse_execution_backend.main \
  --repo-url https://github.com/your-org/your-repository.git \
  --task "Locate the failing test, fix the implementation, and run the targeted tests." \
  --max-steps 20

Use --repo-ref for a branch, tag, or commit, and --template when the repository needs an E2B template with preinstalled system dependencies.

Notes

• The model can execute arbitrary commands only in its E2B sandbox. write_file also rejects paths outside the repository workspace.
• Command output is capped before it is returned to the model.
• This is a reference cookbook, not a new SDK. It currently supports public HTTPS repositories and one persistent sandbox per run.

Test

The tests replace E2B and the Meta API client with local test doubles, so they do not create a sandbox or make an API request.

uv run pytest
uv run python -m compileall src tests

A live end-to-end run creates an E2B sandbox and calls the Meta Model API; it requires explicit approval and a disposable repository.