Muse Code: Meta’s terminal coding agent

Meta released Muse Code in beta on 5 August 2026, and if you already live inside Claude Code or Codex CLI the question is a narrow one: what is here that you do not already have. This article goes through what Meta claims, what its own charts say, what a session actually costs, and what is still missing from the beta.

Key takeaways

• It comes from Meta Superintelligence Labs and runs on Muse Spark 1.2, a proprietary model served only from Meta’s cloud.
• Meta trained the model together with the agent rather than separately, and says it used trajectories from the harness itself as training data.
• The subagents do not spawn and die per task: they stay alive for the whole session.
• Every model call, tool run, approval and edit is appended to a local event log, and both the audit trail and the crash recovery come from it.
• The price listed on 30 August 2026 is $1.25 per million input tokens and $4.25 output, with a second, far cheaper tier you pay for with your data.

What Muse Code is and who built it

Muse Code is a coding agent that lives in the terminal. The announcement’s author metadata and its JSON-LD both attribute it to Meta Superintelligence Labs, the unit Meta assembled in 2026 to concentrate its model work. The opening line leaves no room for interpretation: "We’re excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model."

The mental fit is easy if you already use Codex CLI or Claude Code: same format, same place, same class of task. Meta positions it for large repositories, planning changes, writing the code and validating the result. Muse Spark 1.2 is available inside Muse Code and through the Meta Model API, with a one-million-token context window.

What co-training the model with the harness means

This is the differentiating claim and it deserves a slow read. Meta does not say it took a good model and built a harness on top; it says it trained both at once. The text describes rejection-sampled harness trajectories and recipe optimisations for goals, context compaction and subagents, plus integrating the Muse Code toolset into training to maximise harness compatibility.

Does it matter? Yes, though not as a sales argument. It matters as a measurement bias. A model trained against one specific harness will do better in that harness than a model meeting it for the first time, and Meta measures its rivals inside precisely that evaluation framework. To its credit, Meta admits this in writing in the methodology PDF: "We note that our evaluation setup (e.g., agent tools and system prompts) may not be specifically tuned for proprietary third-party models. Therefore, the results may not reflect these models’ best performance when used in environments tailored to their specific strengths."

That warning makes the result more interesting, not less. Meta plays at home, with a model trained for the home ground, and its own charts still put Opus 5 ahead on all three benchmarks it publishes.

The numbers Meta publishes

Something worth saying before the table: the figures do not appear as text in the announcement. They exist only as labels inside three PNG bar charts, which explains why press coverage has reproduced them inconsistently. These are the values read straight off those images, transcribed in full without trimming the list of models.

The methodology PDF supplies the context the bars hide. Terminal-Bench 2.1 uses 89 tasks and reports mean pass@1 over five attempts, each in an isolated container. DeepSWE v1.1 is 113 tasks across 91 repositories and five languages, also at five attempts. Meta’s internal bench gathers 440 tasks derived from real pull requests against its own code, with internet access disabled and two attempts per task. Reasoning effort is not uniform either: xhigh for Meta’s models, max for Opus and GPT, high for Grok and Gemini.

It is worth stressing that this column is Meta’s own internal run, not the official Terminal-Bench board. They do not agree. The board the benchmark itself publishes has no Opus 5 row at all on version 2.1, and its top two places are Claude Code with Fable 5 at 83.82% and Codex with GPT-5.5 at 83.15%. Version 2.1 also shipped on 6 May 2026 and has been superseded by 4.0, from 28 August, where the picture changes completely. So the 86.7% above is a Meta measurement using Meta’s own scaffolding, not a figure from the public board, and it should be read that way.

Against its predecessor the gain is clear and checkable: Muse Spark 1.1 with mini-swe-agent scored 76.2% on Terminal-Bench 2.1, so 1.2 inside Muse Code adds 6.7 points. Against the competition, the honest reading is that it gets close without arriving. This table is Meta’s, not ours: the three-way comparison against the official board is at Claude Code vs Codex CLI vs Muse Code in 2026.

Subagents that never restart

The architecture Meta describes is a simple agent loop plus a set of async background agents. The difference from what you already know is in the word persistent: those background agents stay alive for the whole session instead of being spawned for individual tasks, which Meta says avoids repeating information gathering and cuts both latency and the need to steer them by hand on long jobs.

On parallel work, TechCrunch reported that the agent fans work out to subagents in isolated Git worktrees when the project is large enough, so your working copy is never touched. Mark Zuckerberg said at the launch that in testing they had it build six features of a game at once with no collisions. That is a demonstration rather than a measurement, and it is worth treating as one.

The agent also ships three skills by default: /plan turns a task into a plan that needs approval, /grill stress-tests that plan until it holds up, and /goal works toward the stated objective.

The event log as audit trail and as rollback

This is the part of the design I find most interesting. Meta describes it this way: the agent uses a local event log to which every model call, every tool run, every approval and every edit is appended. Being the single source of truth, the runtime becomes replay-exact and restart-safe; after a crash the agent resumes exactly where it stopped.

In practice that gives you two separate things worth not conflating. The first is an auditable trace: you can reconstruct why the agent touched a file and under which prior approval, which is exactly what is missing for anyone who has tried to justify a machine-generated change to a colleague. The second is the ability to pick a many-hour task back up without starting over, and that is what makes Meta’s case study viable: optimising GPU kernels across more than a thousand tool calls and up to 24 continuous hours on NVIDIA Hopper cards.

It is worth saying what the log is not. It is not version control and it does not replace Git; it is the agent’s diary, not your repository’s.

Installing on macOS and Linux

The command Meta publishes is a single line:

curl -fsSL https://dev.meta.ai/install.sh | bash

I downloaded that script before writing this, because a pipe into bash deserves reading. What it does is more contained than I feared and less transparent than I would like:

1. It installs a launcher called muse into ~/.local/bin, a path you can change with MUSE_INSTALL_DIR.
2. It appends the PATH line to .zshrc, .bashrc, .profile or conf.d/muse.fish depending on your shell, unless you set MUSE_NO_MODIFY_PATH.
3. It verifies the launcher’s SHA-256 only when the server sends an x-content-sha256 header. The script’s own comment says so: "Verified when advertised; older serving code sends none."
4. It downloads the real binary from lookaside.facebook.com and asks you to sign in through an OIDC device-code flow against auth.meta.com; the credentials land in ~/.config/muse/auth.json.
5. It leaves an update check running every 3,600 seconds against the muse-stable channel, which you turn off with MUSE_NO_AUTO_UPDATE=1.

On 30 August 2026 that channel served version 1.0.1-R1848.1. The manifest for that version declares binaries of 213.0 MiB for Linux on arm64, 238.6 MiB for Linux on x86_64, 217.6 MiB for macOS on Apple Silicon and 237.3 MiB for macOS on Intel. It is a heavy, self-contained binary with no Node or Python dependency.

One detail that appears in no press release: that same manifest also lists x86_windows (277.0 MiB) and aarch64_windows (251.4 MiB) artefacts. They exist. What does not exist is a way to install them, because the launcher’s platform-detection function only recognises macOS and Linux on x86_64 and arm64, and aborts with "unsupported platform" for anything else. Windows users go through WSL, which as far as the installer is concerned is Linux.

What a real task costs

Meta publishes two variants of the same model. These prices were read from the product page on 30 August 2026 and can change:

The gap is 12.5x on input and 21.25x on output. Put into a concrete session, and assuming a million input tokens at an 80% cache hit rate plus 100,000 output tokens, the arithmetic gives roughly $0.80 on the standard tier and roughly $0.04 on the contributor tier. A million-token input session is not a laboratory example: with a one-million context window, a mid-sized repository re-read a few times gets there without effort.

That is cheap, and it is Meta’s thesis. Alexandr Wang, the company’s AI chief, put it to TechCrunch this way: "We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective."

The asterisk sits on the cheap tier. Paying four cents instead of eighty means authorising Meta to train on your prompts and on the completions, which is to say on fragments of your code. For a personal project or for code that is already public, that is a bargain. For a client’s repository, it is not a decision you get to make on your own.

What is still missing in beta

An honest list, unvarnished:

• No open weights. Muse Spark 1.2 is proprietary and served only from Meta’s infrastructure. Asked about it, Zuckerberg said he would have more to share soon, which is not a commitment.
• No Windows install even though the binaries exist in the manifest.
• No public repository or changelog for the agent. Documentation lives in Meta’s developer portal, behind the account door.
• A Meta account and a payment method are required before the first run.
• It is version 1.0.1. Three weeks of public life and no release notes published along the way.
• No editor integration announced. Terminal and nothing else, which for many people is a feature.

Frequently asked questions

Does Muse Code replace Claude Code or Codex CLI?

Not today, if your criterion is model capability. In the three charts Meta itself publishes, Opus 5 is ahead of Muse Spark 1.2. If your criterion is cost per token on context-heavy tasks, the conversation flips and it is worth a trial.

What happens to my code on the cheap tier?

Meta states it plainly on its own product page: on the contributor variant, your prompts and the completions are used to improve the product. On the standard variant, they are not. The price difference is literally the price of your data.

Does the event log let me undo changes?

It lets you reconstruct and resume, which is not the same as revert. It stores every action in order and allows replaying the session and picking it up after a crash, but rolling files back is still Git’s job. Work on a branch, as with any other agent.

Conclusion

Muse Code is not a capability breakthrough: on Meta’s own numbers, measured with a harness Meta trained alongside the model, Opus 5 wins all three benchmarks. What it does bring is two defensible things. One is the runtime design, with an event log that turns an agent session into something auditable and resumable, which is exactly what hours-long tasks need. The other is price, with a competitive standard tier and a cheap tier that only makes sense when your code can be somebody else’s training data.

My practical advice: install it, give it a real refactoring task in a large repository and compare the bill with your current tool’s. Do not make it your default agent yet. If you are arriving from elsewhere, start with the round-up of Claude Code, Cursor and Copilot in 2026 to place the ground. The Spanish version of this article is at Muse Code: el agente de terminal de Meta.

Sources

1. Meta Superintelligence Labs, introducing Muse Code and Muse Spark 1.2
2. Meta, Muse Spark 1.2 and Muse Code evaluation methodology
3. Meta for Developers, Muse Code product page
4. Meta, Muse Code installer script
5. TechCrunch, Meta launches Muse Code
6. VentureBeat, Meta enters the AI coding wars