The headline: On August 5, Meta shipped Muse Code — its first AI coding agent. It runs in the terminal like Claude Code and Codex, is powered by the Muse Spark 1.2 model, and is priced at roughly a third of the market leaders. Meta says it understands large codebases, runs parallel sub-agents in isolated worktrees, and can resume a crashed 24-hour task from its local event log. For teams, the interesting questions are the same as with every coding agent: what does it actually do on your code, where does your data go, and does the price hold at your scale?
What just happened
On August 5, Meta announced Muse Code as a beta, a terminal-based coding agent aimed squarely at Claude Code and OpenAI Codex. Mark Zuckerberg announced it on X, describing it as an agent that can “complete complex software engineering tasks in large codebases, including analyzing projects, planning modifications, writing code, running tools, and verifying results.”
It runs on the Muse Spark 1.2 model, the latest coding-focused update to Meta’s Muse family, and installs with a single terminal command on macOS and Linux. Access is developer-first: API keys via Meta’s developer portal and the OpenRouter hosting platform.
The four features that matter
1. Codebase-aware context. Muse Code is trained to index and understand an entire repository — project architecture, dependencies, and conventions — rather than reasoning file by file. Meta reports it completed tasks like adding end-to-end encryption to Instagram’s backend services and refactoring WhatsApp’s message routing module, work it says would take senior engineers days.
2. Parallel sub-agents in isolated worktrees. Instead of one prompt-and-respond loop, Muse Code coordinates persistent background agents that run concurrently in isolated worktrees. Zuckerberg says the team tested six features built in parallel for one game without any conflicts. The design goal: multiple refactors without the codebase turning into a mess, and long tasks split into parallel subtasks.
3. Crash recovery via a local event log. Every model call, tool run, approval, and edit is written to a local event log that acts as the single source of truth. If the agent crashes, it resumes from the breakpoint instead of restarting. Meta describes a 24-hour run on Nvidia Hopper GPUs involving over 1,000 tool calls — precisely replayable from the log.
4. Built-in skills. /plan produces a gated architecture plan that requires approval before execution; /grill stress-tests a plan until it is robust; /goal drives multi-step execution toward a target. These map to the workflow gates teams already use: plan, review, approve, execute.
The price war
Muse Code is priced at $1.25 per million input tokens and $4.25 per million output tokens — under 40% of Claude Code and Codex, according to Meta’s announcements. A “contributor” subscription tier costs roughly 10× less in exchange for contributing data to model improvement, and new API accounts get $20 in free credits without a card. Meta also advertises a zero-data-retention option.
The pricing is the strategically interesting part. Meta is doing what it does best: undercutting on price, using the scale of its open ecosystem, and trading subsidized access for data. The “contributor tier” is effectively a data-for-discount trade — worth reading the terms before accepting it in a team setting.
The open-source question
Asked on X whether Muse Code would be open-sourced, Zuckerberg replied that there would be “more to share on this topic soon.” That is not a commitment. If Muse Code stays closed, it competes directly with Claude Code and Codex. If it opens up, it changes the math for teams that want to keep agents inside their own infrastructure — the same question the site has covered for open-weight models and self-hosted platforms.
What engineering teams should verify
Treat the launch as a claim sheet, not a spec sheet. Before adopting Muse Code, or any coding agent, verify the same things:
- Acceptance on your code, not benchmarks. Meta’s internal results come from Instagram and WhatsApp scale codebases. Your repository, your tasks, your acceptance criteria — run the same bounded pilot you would run for any agent or model.
- Data path and retention. The zero-data-retention option needs to be confirmed against how the agent routes context, logs, and telemetry. The local event log is local; confirm what leaves the machine.
- Total cost at your scale. $1.25/$4.25 per million is the sticker. Agent tasks burn many tokens per task, and long-context routing compounds the bill. Model the real per-task economics before comparing with alternatives.
- Review and approval gates. The
/planskill maps to team review gates, but only if the gates are actually enforced in the workflow. A gated plan you never read is not a control.
MonkeyCode’s positioning is different by design: instead of a single terminal agent, it runs managed agent work in a shared, self-hostable environment, with a free tier of 30M tokens per day for evaluation. Teams evaluating Muse Code should weigh per-seat terminal pricing against a shared platform where the environment — and the data path — is one you control.
The take
Meta entering the coding-agent market is significant for three reasons: it legitimizes the terminal-agent category as the battleground, it pressures the price structure of every competitor, and its answer to the open-source question — still pending — will decide whether this becomes another closed subscription or a new open ecosystem. For engineering teams, none of that changes the evaluation checklist: prove it on your code, verify the data path, and price it on real per-task economics.