Meta has shipped Muse Code, a coding agent built to run unattended in the terminal and inside continuous-integration pipelines, not only alongside a developer at a keyboard. It runs on Muse Spark 1.2, a coding-focused model Meta updated alongside it. The pitch is unusual for this category: Meta leads with how thoroughly a session can be reconstructed after the fact, not with a leaderboard number.

Rival coding agents from OpenAI, Anthropic and Google have spent much of 2026 competing mostly on capability claims: which model writes cleaner code, clears more of a benchmark suite, or needs fewer retries. Muse Code carries a benchmark score too. Meta’s own testing put the model at 82.9 percent on the Terminal-Bench 2.1 suite. But the documentation spends far more space on approvals, sandboxing and audit trails than on that figure.

Setup is a single install script that places a command-line binary on the user’s machine. Interactive use authenticates through a browser; CI pipelines and other headless environments instead use an API key. Approvals and an operating-system-level sandbox are enabled by default, and Meta’s own guidance is to start with small, easily checked tasks rather than open-ended ones.

Under the hood, one agent drafts a change, a second critiques it, and tests execute automatically. When a task can be broken into parts, the work splits across several persistent background agents running at once. Meta says each one runs inside its own git worktree, which keeps parallel edits from overwriting each other while the main project directory stays untouched.

Every agent spawned, every tool invocation, and every time a human steers or cancels a run gets appended to a JSON Lines log stored locally. That log is what makes a session replayable step by step and recoverable if the process crashes mid-task, according to Meta. A separate /effort command lets a user cap how much reasoning compute a task burns, since thinking tokens are billed as output.

That emphasis is the real story here. In a market where every lab claims its model is the smartest coder, Meta is selling the ability to prove what an agent did and why, down to the individual tool call. That matters most the moment a team lets these agents run inside CI with nobody watching each step. A wrong benchmark claim costs a lab credibility. An unauditable agent that silently rewrites a teammate’s branch overnight costs an engineering team actual hours.

The agent arrives with four skills already wired in:

Meta describes three gains in Muse Spark 1.2: closer alignment between the model and the Muse Code harness without losing usability in other coding agents, stronger results on tasks spanning an entire repository, and a context window Meta lists at one million tokens. The model is reachable through Muse Code itself, Meta’s Model API, or OpenRouter. Meta says its API can substitute directly for the OpenAI and Anthropic SDKs, or for command-line tools such as OpenCode and Claude Code.

Pricing splits into two tiers. A contributor tier limits usage by token volume across rolling five-hour windows in a set of eligible countries. Meta lists standard pricing at $1.25 for a million input tokens, $4.25 for a million output tokens, and $0.15 when the input is served from cache. Teams that need zero data retention can request it through Meta’s sales organization.

Muse Code remains in beta. Muse Spark 1.2 reached wider global availability on August 5. For teams already running coding agents unattended in CI, the test worth running this quarter is not another benchmark comparison. It is whether Muse Code’s event logs would actually let an engineer reconstruct what an agent did after a bad merge, and whether that assurance is worth switching stacks for.

Per Meta’s AI Developers blog and its Muse Code documentation, published August 5, 2026.