Meta has released Muse Code (beta), a terminal-based coding agent powered by Muse Spark 1.2, the company's updated coding-focused model. The tool is now available through the Meta Model API and a curl install script.

Muse Code coordinates persistent background subagents and maintains a replay-exact event log, enabling the agent to resume precisely where it left off after a crash.

On Meta's own benchmark charts, Muse Spark 1.2 trails Anthropic's Opus 5 across every coding benchmark shown, while beating OpenAI's Codex and Google's Antigravity on most metrics.

Meta is the latest tech giant to ship a dedicated coding agent, joining a competitive field dominated by Anthropic and OpenAI. The release signals that terminal-native coding agents—popularized by Anthropic's Claude Code—have become a standard competitive frontier, with each major lab now shipping or developing its own equivalent.

"We're excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model," the company wrote in an official announcement. "This marks our next step toward the frontier, with larger and much more capable models on the way."

Built for Large-Repository Engineering

Muse Code is designed for software engineering across large repositories. According to Meta, it "takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results. It can coordinate multiple persistent subagents for each task, solving difficult problems faster, more accurately, and with less intervention."

A standout feature is the runtime architecture. Muse Code logs every model call, tool run, approval, and edit to a local event log that serves as a single source of truth.

"This single source of truth makes the runtime replay-exact and restart-safe: after a crash, the agent can resume precisely where it stopped," Meta said. For long-running engineering jobs, this capability may matter more than raw processing speed—a dimension that competing products have not emphasized. Crash resilience is a notably underaddressed problem in agentic coding, where multi-hour tasks can be derailed by a single runtime failure.

The agent ships with several default skills. The "/plan" command converts a task into an approval-gated plan, "/grill" stress-tests that plan until it holds up, and "/goal" works toward successful completion of the objective, similar to what Hermes does. Meta said it co-trained Muse Spark 1.2 alongside Muse Code so that the core LLM and the agent operate in synergy.

Benchmarks Show Progress, but Not Supremacy

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. Meta said it "significantly scaled up training compute on coding tasks while expanding training environment diversity, delivering improvements in code generation, complex debugging, and end-to-end developer workflows."

On Terminal-Bench 2.1, Muse Spark 1.2 with Muse Code scored 82.9%, behind Claude Code on Opus 5 at 86.7% but ahead of GPT-5.6 Terra on Codex (81.8%) and Grok Build (81.6%).

DeepSWE 1.1, which measures agentic coding capabilities, showed a narrower gap: 59.3% for Muse versus 65.0% for Opus 5 and 64.8% for Codex. On Meta's internal coding bench, Muse reached 70.6% compared to Opus 5's 79.4%.

The speedup charts reversed the ordering. Over 1,000-plus tool calls, Opus 5 posted the biggest gain versus baseline (approximately 74–75%), with Muse Spark 1.2 in the mid-range at roughly 61–69% depending on the run. Meta emphasized that the agent continues improving as tool calls accumulate—a desirable behavior for a long-horizon coding system.

Long-Horizon and Multimodal Demos

The most notable demonstrations focused on long-horizon and multimodal capabilities. During stress testing, Meta said Muse Code "iteratively optimized GPU kernels over 1,000+ tool calls (up to 24 hours) on Nvidia Hopper GPUs," demonstrating the ability to improve performance over extended periods.

The tool also offers visual-coding functionality. In one demo, a user drops a fly-through video of a house into the terminal as an mp4 file, and Muse Code "interprets the video and produces a visually rich website with booking capabilities." The ability to convert raw video into a working web application represents the multimodal pitch Meta has been advancing across the Muse product line.

See the launch thread:

Here's an example of Muse Code's multimodal visual coding capabilities. In this demo, the user inputs a fly-through video of a home into the terminal as an mp4 file. Muse Code interprets the video and produces a visually rich website with booking capabilities. pic.twitter.com/3CAfIMYmAB

— AI at Meta (@AIatMeta) August 5, 2026

A Crowded Field

Meta enters a market with established players. OpenAI's Codex already runs parallel cloud agents; DeepSeek has built its own rival to Claude Code; and agentic tools like Hermes and OpenClaw are already functional substitutes with expanding capabilities.

Muse Code's distinguishing features are its crash-safe runtime and subagent coordination architecture rather than benchmark leadership. For Meta, the play appears to be reliability and autonomy over raw model dominance—a bet that enterprise and research developers will prioritize dependable long-running agents over marginal benchmark gains.

The inherent risk of agentic coding applies: an agent that resumes after a crash and continues calling tools for 24 hours is both powerful and unpredictable. Meta is betting that developers want that degree of autonomy.

Muse Code is available for testing upon installation by entering the following command:

curl -fsSL | bash