On August 5, 2026, Meta unveiled Muse Code, its own AI coding agent, putting it in direct competition with Anthropic’s Claude Code and OpenAI’s Codex. The tool runs in the terminal, independently plans and writes code for larger repositories, and is currently available as a beta for macOS and Linux. Muse Code is powered by the new Muse Spark 1.2 model, which Meta released at the same time.
In a nutshell:
- Muse Code is a terminal agent for complex software development, currently in beta for macOS and Linux.
- The model behind it is called Muse Spark 1.2 and is also accessible via the Meta Model API.
- Installation is done with a single terminal command, and access is available via dev.meta.ai.
- Meta has not yet announced pricing, but is positioning the tool as a more affordable alternative to Claude Code and Codex.
What is Muse Code?
Muse Code is installed via the command line and is also launched from there. Instead of writing individual lines of code, the agent handles entire software development tasks in large repositories: it plans changes, writes the code, and then checks whether the result works. Meta is clearly targeting the same audience as Anthropic’s Claude Code and OpenAI’s Codex, but with a cost advantage—more on that below.
If you want to give it a try, you can install Muse Code with a single command in the terminal:
curl -fsSL https://dev.meta.ai/install.sh | bashMeta has not yet announced a Windows version. If you’re using Windows, the tool is currently only available via WSL.
Muse Spark 1.2: The Model Behind the Agent
Behind Muse Code lies Muse Spark 1.2, a language model specialized for programming tasks. Meta trained it specifically for long coding sessions—that is, for working across multiple files and repositories—rather than for short autocomplete suggestions. One detail about the training process is interesting: the previous version, Muse Spark 1.1, generated part of the training data for 1.2 itself—essentially a self-improvement loop. Additionally, the model and the agent were trained in parallel to ensure they work together as effectively as possible.
Via Meta’s Model API, Muse Spark 1.2 can now also be used outside of Muse Code, with expanded global access, according to Meta. The model weights are not available for download; the model runs exclusively on Meta’s servers.
Here are the features Muse Code offers
A few technical details set Muse Code apart from traditional autocomplete tools:
Async Background Agents: Instead of starting a new process for each subtask, Muse Code keeps a set of background agents active throughout the entire session. This is intended to prevent duplicate data retrieval and unnecessary wait times during multi-step tasks.
Event Log Runtime: Every model call, every tool execution, and every edit is recorded in a local log that is only appended to and never overwritten. It sounds like bookkeeping, but it’s practical: If a session crashes, it can be resumed at exactly the same point without losing any context.
Bundled Skills: Three built-in commands control the workflow. /plan converts a task into a plan requiring approval, /grill stress-tests this plan before it is implemented, and /goal keeps the agent focused on a defined goal.
Parallel Sub-Agents in Isolated Worktrees: For larger projects, Muse Code allows multiple sub-agents to work simultaneously, each in its own workspace that is separate from the main directory. According to a TechCrunch report, Mark Zuckerberg used the tool himself to have six features for a game developed simultaneously without the changes overwriting one another.
Benchmarks and a 24-Hour GPU Case Study
Meta tested Muse Spark 1.2 against three benchmarks:
| Benchmark | Scope |
|---|---|
| Terminal Bench 2.1 | 89 tasks, measured as Pass@1 over five runs |
| DeepSWE v1.1 | 113 tasks from 91 repositories in 5 programming languages |
| Meta Internal Coding Bench | 440 tasks from real pull requests |
So far, Meta has only published specific percentages in bar charts, without providing exact figures in the body of the text. A case study on GPU kernel optimization, described in the official blog post, provides more meaningful insight: Over more than 1,000 tool invocations, spread out over up to 24 hours, the model optimized kernels for NVIDIA’s Hopper architecture. For so-called KDA kernels, it combined chunk-parallel preparation with a sequential inter-chunk scan and, according to Meta, achieved noticeable improvements over the specified baseline. Anyone who has ever tried to tune GPU code by hand knows how much fiddly work is usually involved; 24 hours straight without a break sounds like a task no one would voluntarily take on. Just as with classic hardware tests like Geekbench, it often takes a few months of real-world use to see how relevant such synthetic benchmarks ultimately are to a developer’s day-to-day work.
Availability, Price, and Installation
| Developers | Meta (Meta Superintelligence Labs) |
| Model | Muse Spark 1.2 |
| Platforms | macOS, Linux (Beta) |
| Installation | curl -fsSL https://dev.meta.ai/install.sh | bash |
| Access | dev.meta.ai |
| Price | Not yet published |
| Release Date | August 5, 2026 |
Meta has so far remained tight-lipped about pricing. Neither a subscription model nor API pricing per token has been publicly disclosed. However, Meta’s head of AI, Alexandr Wang, has already hinted to TechCrunch about the direction the company is heading: Muse Code could be “an incredibly good option,” “especially from a cost perspective.” Meta is thus positioning the tool quite openly as the more affordable alternative to Claude Code and Codex, though specific figures are still lacking.
Muse Code, Claude Code, and Codex: A Comparison
With Muse Code, Meta is joining a field that is currently evolving rapidly. Anthropic has taken the lead with Claude Code, OpenAI is following suit with Codex, and now Meta is joining the fray with its own model and its own terminal agent. It’s striking how heavily Meta is focusing on fault tolerance and parallelization in its architecture; the event log and isolated worktrees for sub-agents seem like a direct response to a problem familiar to experienced users of coding agents: interrupted sessions and conflicting changes during parallel tasks.
Whether Muse Code can keep up with Claude Code or Codex in everyday use cannot yet be seriously assessed after just a few days in beta. The three benchmarks mentioned are Meta’s own selection; independent comparison tests are still pending. It will be particularly interesting to see what becomes of the announced cost-advantage strategy once Meta announces specific prices.
Conclusion
Muse Code shows that Meta no longer intends to leave the market for AI coding agents solely to Anthropic and OpenAI. Technically, the tool offers some well-thought-out features, including persistent background agents, a restart-safe event log, and parallel sub-agents in isolated worktrees. Anyone who wants to start experimenting on macOS or Linux right now can install the beta via a terminal command at dev.meta.ai. However, a final assessment will require independent testing in real-world developer workflows, and—most importantly—concrete pricing details still need to be announced.
Sources: Meta AI Developer Blog, Meta AI Research, TechCrunch, 9to5Mac, MarkTechPost
