Meta ships Muse Code, its first terminal coding agent, on Spark 1.2
The macOS/Linux agent keeps background agents alive all session on a 1M-token model, and a contributor tier drops input to $0.10/M if you let Meta train on you.
Copy markdown
Meta's first coding agent, in your terminal
Muse Code is a beta terminal agent for macOS and Linux that plans changes, writes code, and validates results across large repos. Install is one line: curl -fsSL https://dev.meta.ai/install.sh | bash. It's Meta's direct shot at Claude Code and Codex.
Agents that stay alive all session
Instead of spawning a fresh sub-agent per task, Muse Code keeps async background agents running through the whole session, backed by a local event log for replay-exact, restart-safe runs. Bundled skills /plan, /grill, and /goal drive the workflow.
Muse Spark 1.2: a 1M-token coding model
The engine is Muse Spark 1.2, a coding-focused refresh with a 1M-token context window and a self-improvement loop trained on whole-repo generation. Meta cites a kernel-optimization run sustained over 1,000+ tool calls.
The pricing edge: pay less if you share
Standard API rates are $1.25/M input and $4.25/M output. A 'contributor' tier that lets Meta train on your prompts drops that to $0.10/M input and $0.20/M output — 12x cheaper in, 21x cheaper out — with cached input at $0.002/M.
How to wire it in today
Muse Spark 1.2 is in public preview through the Meta Model API and on OpenRouter, so you can drop it into existing tools now. Meta shows benchmarks on Terminal-Bench 2.1, DeepSWE 1.1, and GDPval, but published them as charts without numbers.