Meta ships Muse Code, its first terminal coding agent
A single curl installs Meta's agent on macOS or Linux; Muse Spark 1.2 runs far cheaper than Claude Opus 5. Plus: CopilotKit's new Slack agent SDK.
Copy markdown
One curl, and you have a coding agent
Install with `curl -fsSL https://dev.meta.ai/install.sh | bash` on macOS or Linux — there's no app, it lives in your terminal. It plans, writes, and validates changes across large repos, with async background agents that keep running through a session.
Muse Spark 1.2: 1M context, priced to move
The model behind it carries a 1,048,576-token context and 131K max output, API-only. Standard pricing is $1.25 in / $4.25 out per million tokens (cached input $0.15) — a fraction of Claude Opus 5 or GPT-5.6 Sol.
The benchmark reality check
Meta reports 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, but independent Vals testing puts it 5th of 45 (71.9%), behind Claude Opus 5 and GPT-5.6 Sol. Its real edge is cost-per-test (~$0.69), not top accuracy.
Crash-safe by design, built for long runs
A local event log makes every session replay-exact and restart-safe, so a crash mid-refactor doesn't lose the run. Bundled skills — /plan, /grill to stress-test a plan, and /goal — steer long-horizon tasks.
The cheap tier costs you your data
A Contributor tier drops pricing to $0.10 in / $0.20 out per million tokens, but Meta may train on your prompts and code; the Standard tier doesn't. Decide per project before you point it at a private repo.
Elsewhere: drop any agent into Slack
CopilotKit open-sourced the MIT-licensed Channels SDK (v0.5.0): route any AG-UI agent — LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK — into Slack (Teams in early access) with native Block Kit UI and human approval gates, not text blobs. Needs Node 22+ and a long-running process.