Muse Code

• Fanning a batch of bug fixes and small features out to parallel subagents, each in its own Git worktree, then reviewing each branch
• Running `muse exec` in CI to run the test suite and summarize failures, with the sandbox kept on and approvals disabled
• Auditing a service with a workflow that has separate agents check API design, security boundaries, and test coverage, then merge one report
• Using `/plan` to get a plan grounded in the actual code and approving it before any edits are made
• Moving from Claude Code or Codex by importing existing skills, memory notes, and MCP servers
• Building an internal tool or editor integration on top of the Muse Session Protocol and TypeScript SDK

Muse Code is a full-featured first-party agent rather than a thin wrapper around Meta's API. The defaults are conservative in useful ways: a real OS sandbox, network destinations approved one at a time, shell commands reviewed stage by stage, and a local event log of everything the agent did. Multi-agent work is central, not an add-on. The appeal is mostly price: $5 a month to start, or Contributor token rates that are a fraction of Standard, as long as you accept Meta training on that usage and a much lower rate limit. The cost is lock-in to Muse Spark, a closed-source CLI, no documented mainstream IDE extension, and features that still vary by platform and rollout. It is a reasonable second agent to try alongside Claude Code or Codex, especially because it reads their instruction files and skills.

About Muse Code

Muse Code is the first-party agent for Meta Model API: you install a native `muse` binary with a one-line install script (a shell script on macOS and Linux, PowerShell on Windows), run `muse` in a project, and sign in with a browser or a Meta API key. The same binary runs an interactive terminal UI or headless with `muse exec` for scripts and CI. Meta built it around Muse Spark and says the model and the agent were trained together. Two things set it apart. First, guardrails are on from the first run. Every shell command runs inside an OS-enforced sandbox (Seatbelt on macOS, bubblewrap on Linux, a Windows sandbox on Windows) that keeps the filesystem outside the workspace read-only and approves network destinations one by one. Compound commands are reviewed stage by stage, and in the default Auto-review profile a model-based approval judge clears requests it considers safe and sends the rest to you. Second, it leans on multi-agent work. A lead agent can spawn subagents, each optionally in its own Git worktree (eight agents per tree at once by default, including the lead), workflows can run up to 1,000 child tasks over a run (at most 16 active at once locally), and four background observer agents handle memory recall, skill recall, goal tracking, and verification. Every model call, tool call, and approval is written to an append-only event log, which powers crash recovery, rewind, forks, and exported trajectories. It reads AGENTS.md and CLAUDE.md, discovers Claude Code and Codex skills, and supports hooks and MCP servers. Pricing is either a Muse Code subscription ($5, $15, or $50 a month, usable only in the CLI) or Meta Model API pay-as-you-go, where Standard models cost $1.25 input and $4.25 output per million tokens and the Contributor models cost $0.10 and $0.20 in exchange for letting Meta train on your prompts and completions (with a much lower rate limit). Meta launched it in beta on August 5, 2026 and took it out of beta on August 31. The tradeoffs: it is built for Meta's own models, the CLI is not open source, no VS Code or JetBrains extension is documented, and some features, such as workflows on the current Apple silicon build and session messaging on Windows, are not available everywhere.