Muse Code review, pricing and limits

For developers who already live in a terminal, Muse Code is a beta coding agent alternative to Claude Code or Codex: same idea of autonomous, multi-step coding, backed by a 1M-token context window and Meta's own Contributor-tier data trade-off for cheaper tokens. It is not yet a safe pick for regulated or spend-capped teams.

Muse Code is Meta's terminal-based multi-agent coding tool, launched in public beta on August 5, 2026 by Meta Superintelligence Labs. It plans, writes, and validates code across large repositories by fanning work out to parallel sub-agents, each running in its own isolated Git worktree with a replay-exact event log.

Maker: Meta · Autonomy: semi autonomous · Maturity: BETA

Underlying models: Muse Spark 1.3, Muse Spark 1.2

About Muse Code

Muse Code is Meta's terminal-based coding agent, built by Meta Superintelligence Labs. Meta introduced it in public beta the week of August 5, 2026, positioning it as an answer to Claude Code and OpenAI Codex. Instead of one model working through a codebase step by step, Muse Code splits a job across a parent process and multiple isolated worker agents that never touch the same files at once (the mechanics are in Key Features).

The agent runs on Muse Spark, Meta's coding-tuned model family, and moved from its 1.2 launch version to 1.3 the following month (see Key Features for what changed). The CLI keeps a full session log of every action, letting a run be replayed and restarted after a failure, and it defaults to a locked-down local sandbox designed to stop work rather than run unprotected if containment cannot be confirmed. The newer model is also trained to ask clarifying questions on ambiguous prompts and confirm before taking consequential actions, which keeps it short of fully autonomous.

Muse Code is built for developers already comfortable in a terminal: side-project and solo developers weighing Meta's Contributor-tier data trade-off against cheaper tokens, and DevOps or platform teams automating CI/CD pipelines that value the sandboxed, containment-first execution model. It also supports multimodal input beyond plain text (see Key Features). It is a weaker fit for enterprise teams that need a hard spend cap or compliance attestations, or for anyone who wants an IDE-native workflow like Cursor or Windsurf instead of a terminal-only agent; none of those attestations are published for Muse Code as of its beta launch.

Access starts with a one-line CLI install, curl -fsSL https://dev.meta.ai/install.sh | bash, on macOS or Linux; there is no native Windows build or official IDE plugin. Pricing detail (subscription tiers and the underlying API's per-token rates) is covered in the FAQ below and in the pricing notes; the short version is that the API side has no flat spend cap, a real difference from most of the coding agents in Meta's weight class.

Muse Code has shipped two notable milestones since its public beta launch: the move to Muse Spark 1.3 and Meta's promise of a max reasoning variant still going through additional safety testing. On one of Meta's own published benchmark charts, Muse Spark 1.2 trailed Claude Code's independently verified score (exact numbers are in Cons and the alternatives FAQ below), and Meta has not yet published a verified score for the newer 1.3 release.

Pricing

Muse Code ships three monthly plans: Everyday Usage at $5/month (10-50 prompts every 5 hours, voice mode), High Usage at $15/month (5x more usage), and Power Usage at $50/month (20x more usage, early feature access). The underlying Muse Spark API also runs pay-as-you-go: the Contributor tier is $0.10 per million input tokens and $0.20 per million output tokens (Meta may train on this data), while the Standard tier is $1.25 per million input and $4.25 per million output (data is not used for training). There is no flat spend cap on the metered API tiers, and the Contributor tier is restricted to US accounts.

Key Features

• Multi-agent by default: A parent agent automatically fans a large job out to write-capable child sub-agents, each working in its own isolated Git worktree, so parallel workers never collide on the same files.
• OS-enforced sandbox: Seatbelt on macOS and bubblewrap on Linux run by default and are designed to fail closed: the agent refuses to execute rather than run unsandboxed if it cannot confirm the sandbox is active.
• Replay-exact event log: Every model call, tool run, approval, and edit is appended to a plain JSONL log, letting a session be reconstructed exactly and restarted after a crash.
• Muse Spark 1.3 efficiency gain: Muse Spark 1.3, which began rolling out to Muse Code on September 2, 2026, needs about 20% fewer tool calls and 25% fewer tokens than the version it replaced, per Meta's own engineering comparison.
• Voice mode and multimodal input: Muse Code accepts image and video uploads alongside text prompts and offers a voice mode for describing a coding task instead of typing it out.
• Two-tier token pricing: A cheaper Contributor tier lets Meta train on submitted code in exchange for a lower per-token rate, while a pricier Standard tier keeps code out of training data; both share the same context window size.

Strengths

• The isolated-worktree-per-subagent design and the replay-exact JSONL event log make Muse Code's multi-agent runs genuinely auditable, which reviewers single out as the most differentiated part of the product compared with single-thread coding agents.
• The Contributor tier's data-for-discount trade is an option most rival coding agents don't offer at all, which makes Muse Code the cheaper choice for teams that don't mind Meta training on their code.
• Meta shipped a model upgrade for Muse Code inside its first month in beta, faster than most competing coding-agent update cycles, and says the new version needs meaningfully fewer tool calls and tokens for the same task (see Key Features).

Weaknesses

• On Meta's self-reported Terminal-Bench 2.1 chart, Muse Spark 1.2 scores 82.9%, versus Claude Code's independently verified 86.7%, and Meta has not published an independently verified score for either Muse Spark version as of September 2026.
• The metered API tiers carry no flat spend cap, unlike most rival coding agents, which ship either a flat monthly plan or a settable spend limit; reviewers have flagged this as a real risk of an unexpectedly large bill.
• The discounted Contributor tier is restricted to US accounts, and neither tier publishes a SOC 2, HIPAA, or data-residency attestation, which rules Muse Code out for regulated or compliance-heavy teams for now.

Visit Muse Code Official Site