/plan — a plan you approve
/plan turns a request into an explicit plan and waits for approval before Muse Code touches files — the gate between describing a task and editing code.
Independent reference
Muse Code is Meta's terminal coding agent, released in beta on 5 August 2026 by Meta Superintelligence Labs and driven by Muse Spark 1.2. What is Muse Code offering that other CLI agents are not? Persistent background agents, a replayable event log, and per-token billing with no subscription.
What is Muse Code in one paragraph, plus the search names it answers to.
Muse Code is a coding agent living in your terminal. You run muse from a repository, describe a task, and Muse Code reads files, edits them, runs commands and iterates. No IDE, no browser app, no desktop client.
One product answers to many search names: Muse Code meta, meta code, meta ai code, meta coding agent, meta coding ai, meta ai coding agent and meta terminal coding agent all point at this binary. Muse Code beta is its release state.
One disambiguation: Meta Spark was Meta's retired augmented-reality platform, unrelated to Muse Spark 1.2 or Muse Code. Adobe Muse and MuseScore share four letters, nothing more.
What is Muse Code architecturally: a tool-calling loop, resident background agents, a local log.
At base sits a familiar loop: Muse Spark 1.2 gets your instruction plus repository context, picks a tool call, and results feed back. What differs is stamina — Meta AI Research reports Muse Code runs exceeding 1,000 tool calls.
Muse Code background agents are asynchronous and stay alive for a whole session instead of respawning per sub-task. An agent that mapped your build system keeps that understanding — no re-reading the same files every turn.
Every step is appended to a local event log. Ordered and never rewritten, it lets a session replay exactly and a killed session restart from its last event — replay-exact and restart-safe are Meta's words. On the Hacker News launch thread, crash recovery was among few features commenters called interesting.
Muse Code fans work across git worktrees so changes proceed at once without colliding. That log sits on your own disk — privacy consequences on the Muse Code privacy page.
Meta names three default Muse Code skills and publishes no total for what else ships in the box.
/plan turns a request into an explicit plan and waits for approval before Muse Code touches files — the gate between describing a task and editing code.
/grill attacks a plan: unstated assumptions, missing edge cases, steps that will not survive your codebase. Cheapest way to stop Muse Code building the wrong thing.
/goal points Muse Code at a target state and lets it work across many steps — the skill behind Meta's long-horizon runs.
Meta writes that Muse Code includes "several default skills" and names only these three. Community reports suggest around ten, including a taste skill, but that is unconfirmed — weighed on our skills and plugins page.
What is Muse Code demonstrably good at: a multimodal website build and a 24-hour GPU kernel run.
Meta's multimodal demo travelled furthest: a user drops an mp4 fly-through video of a house into a terminal, and Muse Code produces a marketing and booking page for it — video in, handled natively rather than transcribed.
Meta's second showcase runs on NVIDIA Hopper hardware: Muse Code optimised GPU kernels against an FLA Triton KDA baseline, making 1,000+ tool calls over sessions up to 24 hours long. Meta describes substantial improvements but publishes no speedup figure, so treat it as Meta-reported; a Hacker News commenter noted Muse lost to GPT-5.6 Sol there.
Both demos point one way: Muse Code targets tasks measured in hours, not autocomplete. For duller failures, see our troubleshooting guide.
One paragraph on the model, because Muse Spark 1.2 specs and benchmarks have their own page.
Muse Code runs on Muse Spark 1.2, Meta's own model, released the same day and co-trained with Muse Code: 1M-token context, image, video and PDF input, roughly 165 tokens per second in Artificial Analysis testing, which scores it 54 on its Intelligence Index against 51 for Muse Spark 1.1. Parameter count and knowledge cutoff are unpublished. Specs, labelled Meta-reported benchmarks and open weights live on the Muse Spark 1.2 page; see also our Muse Spark API guide.
Muse Code supports macOS and Linux only, installs with one command, and signs in through Meta OAuth.
Its launcher hard-codes four targets — arm64 and x86, macOS and Linux — so Muse Code runs nowhere else natively. One command from Meta's install script places a muse binary in ~/.local/bin:
curl -fsSL https://dev.meta.ai/install.sh | bash
Sign-in is a device-style OAuth flow through auth.meta.com: Muse Code prints a verification link and polls until you approve. That is why it fails inside Docker containers, which lack a browser and a usable /dev/tty.
Windows has no native build; WSL2 is the working route, and Meta has said nothing about supporting it — see our Windows guide. Setup sits on the Muse Code install guide.
What is Muse Code costing you? Two per-token tiers, no monthly plan, and no advertised free allowance.
Muse Code has no subscription: attach a payment method and pay per token across two tiers. Standard is $1.25 per million input and $4.25 per million output; Contributor is $0.10 and $0.20 — roughly twenty times cheaper if you let Meta train on your prompts. All as reported, checked against Meta's pricing and rate limits docs. There is no hard spending cap, only email alerts — early users' loudest complaint. Detail sits on the Muse Code pricing page and our cost calculator.
What is Muse Code worth choosing over? Here it sits beside the three agents it meets most.
| Muse Code | Claude Code | OpenAI Codex | Cursor | |
|---|---|---|---|---|
| Form factor | Terminal only | Terminal + desktop app | CLI, IDE and cloud | AI IDE |
| Model | Muse Spark 1.2 | Opus 5 / Sonnet 5 | GPT-5.6 Sol / Terra | Composer 2.x + frontier |
| Entry price | Per token, no plan | Pro $17–20/mo | Go $8 / Plus $20 | Pro $20/mo |
| Native Windows | No | Yes | Unverified | Yes |
| Free tier | None advertised | None | Yes | Yes (Hobby) |
| Standout | 1M context, cheap tokens, video input | Cross-platform, MCP, enterprise controls | Cloud async tasks | IDE-native agent |
Competitor prices as reported, August 2026. Head to head: Muse Code vs Claude Code, Muse Code vs Codex, Muse Code vs Cursor, Muse Code vs GitHub Copilot.
Where each thread in this what is Muse Code overview continues in depth.
Behind on benchmarks, ahead on price. Artificial Analysis measures Muse Spark 1.2 at 80% on Terminal-Bench v2.1, and Meta's own reported 82.9% still trails Claude Opus 5 at 86.7%, as relayed by press. What is Muse Code winning on? Token cost and context length — see our full comparison.
No. Neither Muse Code itself nor Muse Spark 1.2 weights are released. The agent ships as a prebuilt binary, and the model is served only through Meta's API and resellers. Asked about open weights, Mark Zuckerberg said he would have more to share soon, as reported by The Register.
Because a frontier model needs distribution and an agent generates training-grade data. Muse Spark 1.2 was co-trained with Muse Code, making it both product and proving ground for Meta Superintelligence Labs. Contributor pricing states the second motive plainly: a roughly twentyfold discount for training rights.
Unconfirmed, and nobody has shown proof. The rumour spread because parts of the CLI feel familiar, but Muse Code ships as a compiled binary with no published source, so nobody outside Meta has demonstrated shared lineage. Meta has not addressed the question, so treat it as speculation.
Not in beta. Muse Code is terminal-only: no VS Code extension, no JetBrains plugin, no desktop app. Meta has made no public statement about IDE integration, so any timeline you read elsewhere is a guess. If IDE-native agents appeal more, our Cursor comparison covers that difference.
There is no advertised free tier. Billing runs per token from your first call, and a payment method is required. Community reports suggest a $20 starter credit existed during the Muse Spark 1.1 era, but whether it still applies is unverified. Read our pricing breakdown first.
Four gaps stand out: native Windows, any IDE surface, a hard spending cap, and documentation for skills beyond /plan, /grill and /goal. Sign-in also fails inside Docker containers, because that OAuth flow needs a browser. So what is Muse Code today? A one-day-old beta with obvious edges.