Meta Muse Code is the company's new terminal-based AI coding agent, released in beta on August 5, 2026. The tool runs on macOS and Linux and is powered by the new Muse Spark 1.2 model. It takes on complete software engineering tasks across large repositories: planning changes, writing code, and validating results. With this launch Meta steps directly into the market where OpenAI's Codex and Anthropic's Claude Code have been the default choices for many developers.
Why Meta Entered the Coding Agent Race
Meta has been building toward this for months. The company replaced the Llama branding with Muse Spark in April, shipped Muse Spark 1.1 with a paid developer API in July, and in June launched an enterprise agent aimed at customer service teams. Muse Code is the first Meta product aimed squarely at software engineers.
The positioning is deliberate. Alexandr Wang, Meta's AI chief at Meta Superintelligence Labs, told the Wall Street Journal that for many workflows Muse Code can be an incredibly good option, especially from a cost perspective. Mark Zuckerberg announced the beta in a post on X, describing tasks that include planning changes, writing code, and validating the results.
How Muse Code Works
Installation is a single command:
Muse Code runs an agent loop plus a set of async background agents. These specialized agents stay active for the whole session instead of being spawned per task, which avoids redundant information gathering. They decide when to carry out the next step and when to report back to the main agent. Meta says this persistence reduces latency and cuts down the steering needed on difficult, multi-step tasks.
Zuckerberg described the fan-out model in his announcement: when a job is big enough, Muse Code fans out to separate sub-agents working in parallel in isolated worktrees. Your working copy is never touched. During trials, he said, it built six features for a game simultaneously with no collisions.
Technical Breakdown
The runtime keeps a local event log where every model call, tool run, approval, and edit is appended. This log is the single source of truth, and it makes the runtime replay-exact and restart-safe. If the process crashes, the agent resumes exactly where it stopped. That property is what lets Muse Code handle long-running work without being derailed by failures.
Three skills ship with the tool. /plan turns a task into an approval-gated plan. /grill probes that plan until it holds up. /goal keeps the agent working toward a specified objective. The approval gate means a human stays in control of what actually gets executed.
Muse Spark 1.2: The Model Behind It
Muse Spark 1.2 is a coding-focused update to Spark 1.1. Meta scaled up training compute on coding tasks and expanded training environment diversity. The model was co-trained with Muse Code itself, using rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, so the model behaves well inside the agent harness.
Long-horizon training is a core theme. Spark 1.2 was trained on whole-repository generation, large end-to-end projects, and auto-research, with planning to sequence work, goal conditioning to keep direction, and context compaction to hold knowledge across a long session. There is also a self-improvement loop: Spark 1.1 generated challenging coding environments and instruction templates, then Spark 1.2 graded candidate solutions against them to build a scalable training set.
The kernel optimization case study shows the depth. Meta ran the agent on GPU kernel optimization for KDA and MLA kernels on NVIDIA Hopper hardware, letting it write, compile, profile, and improve kernels over more than 1,000 tool calls across up to 24 hours. The models were barred from importing third-party kernel libraries; they had to implement the algorithm in Triton from first principles. The agent kept improving on the baseline over the full run.
Developer Experience
There is no dedicated app interface. Muse Code is a terminal tool, which will appeal to some developers and feel bare to others. The local event log means work survives crashes, and the restart-safe design supports genuinely long tasks. Muse Spark 1.2 is also available through the Meta Model API with expanded global access, so teams can call the model outside the agent.
Muse Code vs Codex vs Claude Code
| Feature | Muse Code | OpenAI Codex | Claude Code |
|---|---|---|---|
| Interface | Terminal CLI | CLI and app | CLI and IDE |
| Install | One-line curl | One-line curl | One-line curl |
| Sub-agents | Persistent async agents | Task-based | Task-based |
| Model | Muse Spark 1.2 | GPT family | Claude family |
| Platforms | macOS, Linux | Cross-platform | Cross-platform |
The table reflects what has been announced so far. Muse Code's differentiators are the persistent background agents, the replay-exact event log, and the explicit cost positioning. Codex and Claude Code ship richer app surfaces and broader platform support today.
Frequently Asked Questions
Is Muse Code free?
Meta has not published pricing. The company is positioning the tool as a strong option from a cost perspective, which suggests it intends to undercut competitors, but no numbers have been announced.
What model does Muse Code use?
Muse Code is powered by Muse Spark 1.2, a coding-focused model that was co-trained with the agent harness. Spark 1.2 is also available through the Meta Model API.
Does Muse Code work on Windows?
The beta supports macOS and Linux only. The install script targets those platforms.
How is Muse Code different from Muse Spark?
Muse Code is the terminal agent that plans, writes, and validates code. Muse Spark is the underlying model. They were built together and ship together.
Can Muse Code work without an app?
Yes, that is the point. It is a terminal tool with no separate app interface, and the local event log keeps it restart-safe after crashes.
Key Takeaways
- Meta now competes directly with OpenAI Codex and Anthropic Claude Code in the coding agent space.
- Persistent async background agents and a replay-exact local event log are the main technical differentiators.
- Muse Spark 1.2 was co-trained with the agent, and the long-horizon training is aimed squarely at whole-repo work.
- Cost positioning could put pressure on pricing across the whole category.
Conclusion
Muse Code is a real entry into the coding agent race, with a few genuinely interesting engineering choices: persistent sub-agents, a crash-safe event log, and a model trained specifically for the harness. The terminal-only beta is available now for macOS and Linux, and the model is open to API users. Whether it pulls developers away from Codex and Claude Code will depend on how the pricing lands.
Sources: Meta Research: Introducing Muse Code and Muse Spark 1.2, TechCrunch: Meta launches Muse Code, 9to5Mac: Meta launches Muse Code AI coding agent for macOS and Linux, Mark Zuckerberg on X
Automated Transmission
This entry was synthesized and populated dynamically using native API integrations.