Audit hundreds of pull requests with a terminal coding agent
Meta launched Muse Code on August 5, 2026. Task: Meta launched Muse Code on August 5, 2026. What it is, how to install it, what Muse Spark 1.2 costs, how the sandbox and subagents work, and whether the benchmarks hold up. Quick answer. Muse Code is Meta's terminal coding agent, launched in beta on August 5, 2026 and powered by the new Muse Spark 1.2 model. It installs with one shell command on macOS and Linux, runs approvals and an OS sandbox by default, spawns parallel subagents into isolated git worktrees, and costs $1.25/$4.25 per million tokens — or $0.10/$0.20 if you let Meta train on your code. Meta shipped its first dedicated coding agent on August 5, 2026. Muse Code is a terminal-native agent in the same category as Claude Code and OpenAI's Codex CLI, built by Meta Superintelligence Labs and running on Muse Spark 1.2, a coding-focused model Meta says it co-trained with the agent itself. This guide is built from Meta's own documentation and the shipped binary rather than the launch-day press cycle — which matters, because a lot of the early coverage is wrong on specifics. We verified the install domain, the command surface, the config paths, and the safety defaults directly. Where Meta has not published something, we say so instead of guessing. What is Muse Code? Muse Code is a single statically-linked binary called muse that you run inside a project directory. It plans a change, edits files, runs shell commands, and checks its own work — the now-familiar agentic loop — but with a few design decisions that genuinely differ from its competitors. The version live at the time of writing is 0.1.0-R708.1, published on the muse-stable channel. It is explicitly a beta. Three things define it: It is local-first. The agent, the subagents, and the session logs all run on your machine. Only the model is hosted. The "persistent background agents" in the marketing are session-scoped and local — widely misread as Meta-hosted infrastructure. It is locked down by default. Approval prompts and an OS-enforced sandbox are both on from the first run, and the sandbox refuses to start if it cannot enforce its boundary. Network access defaults to proxy-only. That is a stricter out-of-the-box posture than most of the field. It is priced aggressively. Two SKUs, a 12–21x gap between them, and the cheap one is paid for with your source code. How do you install Muse Code? One command, macOS or Linux: A warning worth repeating, because several published guides get it wrong: the domain is dev.meta.ai, not dev.ai. The latter is a different site entirely and does not serve the installer. The installer needs only curl, mktemp and bash. There is no Node.js, Python or Homebrew dependency — the binary is statically linked. It installs a self-updating launcher to ~/.local/bin/muse and appends that directory to your shell profile. Supported platforms are macOS and Linux on x86_64 and arm64. Native Windows is not supported and the installer hard-fails on it; WSL2 is the route in. Verify with: For the full walkthrough — auth, CI keys, headless runs and the flags that matter — see our step-by-step Muse Code install guide. What is Muse Spark 1.2? Muse Spark 1.2 is a coding-focused point release of Muse Spark 1.1, not a new base model. Meta describes it as co-trained with Muse Code using rejection-sampled harness trajectories, plus a self-improvement loop in which Spark 1.1 generated training environments and graded candidate solutions for 1.2. It shipped closed-weights — no download, no self-hosting, no fine-tuning. That is changing: on August 10, 2026 Mark Zuckerberg announced Meta will open-source Muse Spark 1.2’s weights, alongside an essay arguing for American open-source AI. The weights are not out yet and no date has been given, so treat it as announced rather than shipped. Meta did ship a genuinely open model that week — see Muse Glimmer below. Published specifications: Context window: 1,048,576 tokens (1M) Max output: 131,072 tokens Default model id: muse-spark-1.2 Reasoning effort levels: none, minimal, low, medium, high, xhigh, ultra — default high Meta has not published the architecture. There is no parameter count, no MoE-versus-dense confirmation, and no stated knowledge cutoff. How do subagents and worktrees work? This is Muse Code's most substantial engineering idea. A lead session spawns child agents, hands each a bounded task, and manages the group while they run.; tools: Meta Muse. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include macOS/Linux, git repository, and API keys. Constraints: No native Windows support, beta stability issues (sign-in bugs), and strict rate limits on cheaper tiers. Human approval is mandatory for all changes. Source reliability is secondary; verify benchmarks independently..
Tools
Published on Agent Work Atlas:
Overview
Meta launched Muse Code on August 5, 2026.
Workflow
Meta launched Muse Code on August 5, 2026. What it is, how to install it, what Muse Spark 1.2 costs, how the sandbox and subagents work, and whether the benchmarks hold up. Quick answer. Muse Code is Meta's terminal coding agent, launched in beta on August 5, 2026 and powered by the new Muse Spark 1.2 model. It installs with one shell command on macOS and Linux, runs approvals and an OS sandbox by default, spawns parallel subagents into isolated git worktrees, and costs $1.25/$4.25 per million tokens — or $0.10/$0.20 if you let Meta train on your code. Meta shipped its first dedicated coding agent on August 5, 2026. Muse Code is a terminal-native agent in the same category as Claude Code and OpenAI's Codex CLI, built by Meta Superintelligence Labs and running on Muse Spark 1.2, a coding-focused model Meta says it co-trained with the agent itself. This guide is built from Meta's own documentation and the shipped binary rather than the launch-day press cycle — which matters, because a lot of the early coverage is wrong on specifics. We verified the install domain, the command surface, the config paths, and the safety defaults directly. Where Meta has not published something, we say so instead of guessing. What is Muse Code? Muse Code is a single statically-linked binary called muse that you run inside a project directory. It plans a change, edits files, runs shell commands, and checks its own work — the now-familiar agentic loop — but with a few design decisions that genuinely differ from its competitors. The version live at the time of writing is 0.1.0-R708.1, published on the muse-stable channel. It is explicitly a beta. Three things define it: It is local-first. The agent, the subagents, and the session logs all run on your machine. Only the model is hosted. The "persistent background agents" in the marketing are session-scoped and local — widely misread as Meta-hosted infrastructure. It is locked down by default. Approval prompts and an OS-enforced sandbox are both on from the first run, and the sandbox refuses to start if it cannot enforce its boundary. Network access defaults to proxy-only. That is a stricter out-of-the-box posture than most of the field. It is priced aggressively. Two SKUs, a 12–21x gap between them, and the cheap one is paid for with your source code. How do you install Muse Code? One command, macOS or Linux: A warning worth repeating, because several published guides get it wrong: the domain is dev.meta.ai, not dev.ai. The latter is a different site entirely and does not serve the installer. The installer needs only curl, mktemp and bash. There is no Node.js, Python or Homebrew dependency — the binary is statically linked. It installs a self-updating launcher to ~/.local/bin/muse and appends that directory to your shell profile. Supported platforms are macOS and Linux on x86_64 and arm64. Native Windows is not supported and the installer hard-fails on it; WSL2 is the route in. Verify with: For the full walkthrough — auth, CI keys, headless runs and the flags that matter — see our step-by-step Muse Code install guide. What is Muse Spark 1.2? Muse Spark 1.2 is a coding-focused point release of Muse Spark 1.1, not a new base model. Meta describes it as co-trained with Muse Code using rejection-sampled harness trajectories, plus a self-improvement loop in which Spark 1.1 generated training environments and graded candidate solutions for 1.2. It shipped closed-weights — no download, no self-hosting, no fine-tuning. That is changing: on August 10, 2026 Mark Zuckerberg announced Meta will open-source Muse Spark 1.2’s weights, alongside an essay arguing for American open-source AI. The weights are not out yet and no date has been given, so treat it as announced rather than shipped. Meta did ship a genuinely open model that week — see Muse Glimmer below. Published specifications: Context window: 1,048,576 tokens (1M) Max output: 131,072 tokens Default model id: muse-spark-1.2 Reasoning effort levels: none, minimal, low, medium, high, xhigh, ultra — default high Meta has not published the architecture. There is no parameter count, no MoE-versus-dense confirmation, and no stated knowledge cutoff. How do subagents and worktrees work? This is Muse Code's most substantial engineering idea. A lead session spawns child agents, hands each a bounded task, and manages the group while they run.
Reported result
Outcomes are unverified for this specific workflow. Independent tests showed mixed results: one reported auditing 222 PRs in under five minutes for ~$0.10, while another failed by hallucinating a non-existent project. No confirmed success metrics exist for large-scale PR audits.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include macOS/Linux, git repository, and API keys. Constraints: No native Windows support, beta stability issues (sign-in bugs), and strict rate limits on cheaper tiers. Human approval is mandatory for all changes. Source reliability is secondary; verify benchmarks independently.
Try it
I want to use Meta Muse Code to audit a batch of pull requests in my repository [replaceable: repo_url]. First, install Muse Code using `curl -fsSL https://dev.meta.ai/install.sh | bash` and verify with `muse --version`. Configure `~/.config/muse/settings.json` to enable the OS sandbox and set approval mode to 'on-request'. Create a list of 5 sample PRs [replaceable: pr_list] for a dry run. Run Muse Code with parallel subagents enabled in isolated git worktrees to generate review comments for these PRs. Do not merge or comment on any PRs automatically. Present the generated review findings to me for human validation. If the dry run is successful and safe, I will provide the full list of PRs [replaceable: full_pr_list] for processing. Ensure no proprietary code is sent to the 'Contributor' tier if confidentiality is required. Measure success by the accuracy of flagged issues versus manual review.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.