Meta ships Muse Code, a terminal coding agent to rival Claude Code
Powered by Muse Spark 1.2, it keeps background agents alive for 24-hour runs — and a contributor tier undercuts rivals ~10x if you share your sessions.
Copy markdown
It plans, writes, and validates across your whole repo
Muse Code is a terminal agent for macOS and Linux, installed with one curl command, that takes complex tasks across large repositories end to end. Bundled /plan, /grill, and /goal skills turn a request into an approval-gated plan before it touches your code — Meta's direct shot at Claude Code, Codex, and Cursor.
Background agents that stay alive all session
Instead of spawning a fresh agent per task, Muse Code keeps specialized background agents running for the whole session to cut latency on multi-step work. Meta says one run made more than 1,000 tool calls over a job lasting up to 24 hours.
A contributor tier ~10x cheaper — if you share sessions
Standard pricing is $1.25 per million input tokens and $4.25 per million output; the opt-in contributor tier is 'more than 10x cheaper' in exchange for letting Meta train on your session data. Meta claims Claude Code's API runs 'substantially higher' for comparable output.
Crash-safe by design
Every model call, tool run, approval, and edit appends to a local event log, so a session is replay-exact and restart-safe after a crash — you resume mid-task instead of starting the whole run over.
Muse Spark 1.2 is in the API too
The model behind the agent also ships in the Meta Model API with expanded global access. Meta benchmarks it on Terminal-Bench 2.1 and DeepSWE 1.1 against Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, and Kimi K3 — but notes its harness may not be tuned for rivals' models, so read the charts with care.