Meta Muse Code is a terminal coding agent for software engineers who run long, multi-file jobs on macOS or Linux. Meta Superintelligence Labs shipped it in public beta on 5 August 2026 alongside Muse Spark 1.2, the reasoning model it was co-trained with. It is Meta's direct answer to Anthropic's Claude Code and OpenAI's Codex.
Core Features
- Four cooperating agent roles - Coordinator, Explorer, Executor and Verifier - that plan, search, edit and validate changes across a whole repository.
- A shared signed event log that survives crashes and restarts, so a session resumes without re-deriving context.
- Git worktree isolation and persistent background agents, with jobs that can run 24 hours or longer instead of the usual 10-15 minute API window.
- A 1M-token context window out of the box, reachable through the Meta Model API and OpenRouter.
- Installs straight into a macOS or Linux terminal; there is no Windows build at launch.
Use Cases / Best For
- Repository-scale refactors and framework migrations that touch dozens of files in a single pass.
- Overnight debugging runs where the agent reproduces a failure, patches it and re-tests on its own.
- Token-heavy teams looking for a cheaper agentic coding loop than Claude Opus 5 or GPT-5.6.
Pricing
Pay-as-you-go at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. A Contributor tier drops that to $0.10 input and $0.20 output, but you agree to let Meta train on your prompts and completions - fine for open-source work, wrong for private repositories holding customer data or secrets. Standard accounts are rate-limited at 3,000 requests per minute.
Our Take
Best for engineers who already live in a terminal and want long-horizon agents at a fraction of frontier pricing. Meta claims 82.9% on Terminal-Bench 2.1 against Claude Opus 5 at 86.7%, but that figure comes from Meta's own harness - Muse Spark 1.1 claimed 80% and later verified at 76.2%, so read the chart as a ceiling. The real trade-off is privacy against price on the Contributor tier. Compare it with Muse Spark 1.1, or browse more AI coding tools.
