Meta Launches Muse Code, a Terminal Coding Agent That Runs Parallel Subagents in Isolated Worktrees

Meta Superintelligence Labs released Muse Code in beta on August 5, 2026, entering the same category as Claude Code and OpenAI’s Codex CLI with a terminal agent designed for large codebases. The tool installs with a single shell command and is built around a coding model, Muse Spark 1.2, that Meta co-trained with the agent itself.

The installation is standard for this type of tool:

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

Available for macOS and Linux.

How It’s Different

Most coding agents spin up a fresh context for each subtask. Muse Code keeps background agents alive for the full session. The persistent architecture reduces the redundant information gathering that happens when agents spawn, acquire context, do a task, and die — only to have the next agent repeat the acquisition step.

Every model call, tool run, approval, and edit is appended to a local event log. If the agent crashes or the session drops, the runtime replays from the log and picks up exactly where it stopped. That makes long-running tasks more viable; you can start an overnight refactor, lose the connection, and come back to a resumed session rather than a lost one.

For large projects, Muse Code splits work across parallel subagents running in isolated git worktrees. Each workaround stays isolated from the user’s working directory, so a subagent generating six features simultaneously doesn’t produce conflicts. Meta tested this and reported no collisions across parallel feature development runs.

The Model

Muse Spark 1.2 is a coding-focused update to the Muse Spark 1.1 model. Meta trained it on long-horizon tasks — whole-repository generation, large project maintenance — using self-improvement loops where the predecessor model generates training data. The release notes specific improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows.

The most striking benchmark case is GPU kernel optimization. Muse Spark 1.2 ran 1,000+ tool calls over 24 hours to optimize NVIDIA Hopper GPU kernels, hitting significant speedup improvements. That’s not a typical developer task, but it’s a useful demonstration that the model can sustain attention across very long runs.

Benchmarks include Terminal-Bench 2.1 and DeepSWE 1.1, with substantial gains over the previous Muse Spark version.

Built-In Skills

The agent ships with three bundled commands:

/plan — creates an approval-gated task plan before doing any work
/grill — stress-tests the plan, poking holes before execution
/goal — works toward a specified objective with lower interruption

The approval-gate on /plan fits the pattern most teams actually need: see the plan, approve or revise it, then let the agent execute. The /grill command is the less common idea — it explicitly tries to break the plan before you commit to it.

Pricing

Muse Spark 1.2 pricing through the Meta Model API:

$1.25 per million input tokens / $4.25 per million output tokens — standard rate
$0.10 per million input tokens / $0.20 per million output tokens — if you allow Meta to train on your code

The training discount is steep. Whether that tradeoff is appropriate depends on what you’re running through it; proprietary business logic is probably not the right fit.

Where Meta Fits in This Race

Claude Code hit 6x growth earlier this year. Codex CLI has been active on GitHub repositories since OpenAI launched it. Cursor owns the IDE space with over $2 billion in ARR before the SpaceX acquisition. Muse Code is entering a market that already has clear leaders, and Meta’s track record on developer tooling has been inconsistent.

What Muse Code has going for it is price, architecture, and Meta’s scale. The persistent agent model and isolated worktree approach are technically sound, and the pricing is aggressive compared to established alternatives. Meta is also distributing through the Meta Model API, which means any team already using the API can add the model without a separate relationship.

The tool is in beta. Access is available now through the Meta Model API and Muse Code’s direct channel.

Sources: