Meta has released a new AI coding tool called Muse Code. The tool is available now in beta for developers to try.
Muse Code is a terminal-based agent. It helps developers write, plan, and check code across large software projects.
The agent runs on Meta's newest model, Muse Spark 1.2. Meta said the model was trained with more focus on coding tasks than earlier versions.
Developers can install Muse Code through the Meta Model API. The company also shared a simple script for setup.
How Muse Code Works
Muse Code can run several background tasks at once. These are called subagents, and each one works on a different part of a coding job.
The tool keeps a detailed log of every action it takes. This includes model calls, tool use, and file edits.
Meta calls this log "replay-exact." If the agent crashes, it can pick up exactly where it left off.
Muse Code also comes with built-in commands. The "/plan" command creates a step-by-step plan that needs approval before it runs.
Another command, "/grill," tests that plan to find weak spots. A third command, "/goal," works to finish the task once the plan is approved.
Meta said the agent can also handle video files. In one demo, it turned a video walkthrough of a house into a working website with booking features.
Benchmark Results
Meta shared test results comparing Muse Code to other coding tools. On the Terminal-Bench 2.1 test, Muse Spark 1.2 scored 82.9 percent.
That score was lower than Anthropic's Opus 5, which reached 86.7 percent. Muse Code did score higher than OpenAI's Codex and Google's Antigravity on this same test.
On a separate test called DeepSWE 1.1, Muse scored 59.3 percent. Opus 5 scored 65.0 percent and Codex scored 64.8 percent on the same test.
Meta also measured how much each tool improved after many tool calls. Opus 5 showed the largest gain, while Muse Code landed in the middle of the group.
Meta said Muse Code kept getting better the longer it ran. In one test, the agent worked on GPU code for up to 24 hours without stopping.
Meta is not the first company to build a coding agent like this. OpenAI's Codex, Google's Antigravity, and Anthropic's Claude Code already offer similar tools.
Other companies, including DeepSeek, have built their own versions too. Meta said Muse Code's main strength is its crash-safe design, not top benchmark scores.
The company said larger and more capable models are planned. For now, Muse Code remains in beta and open for developers to test.
