Muse Code Launch Week: What Is Confirmed and What Is Not

"Documented" means Meta published it. "First-hand" means we read the artefact ourselves. "As reported" means multiple independent third parties agree but no primary source is machine-readable. "Community-reported" means one or more developers claimed it and Meta has not confirmed it. Our full grading rules are on the methodology page.

Almost every Muse Code launch write-up reproduced the same one-line install command and stopped there. The Muse Code installer behind that command is readable, and it settles things people were guessing at: the binary installs to ~/.local/bin, the command is muse rather than muse-code, MUSE_INSTALL_DIR moves the install location, and MUSE_NO_MODIFY_PATH skips the shell PATH edit.

The launcher also pins the release channel to muse-stable and fetches the binary from Meta's CDN with a bearer token. Muse Code sign-in is an OAuth device flow through auth.meta.com that prints a verification link, opens a browser and polls — and it reads from /dev/tty.

That last detail is the whole Docker story. A container with no controlling terminal and no browser cannot complete the flow, which is exactly the failure the Cline team hit publicly. It is not a mysterious bug; it is the documented shape of the Muse Code login. Our install guide covers the full sequence and the troubleshooting page works through it symptom by symptom.

Muse Code has no subscription and no free quota. You attach a payment method to Muse Code and pay per token. The Standard tier is reported at $1.25 per million input tokens and $4.25 per million output; the Contributor tier drops to $0.10 and $0.20 in exchange for granting Meta the right to train on your prompts and completions.

That gap — roughly 12x on input and 21x on output — is the sharpest edge of the Muse Code launch. At those rates the Contributor tier undercuts DeepSeek V4-Flash by about 29%, which makes Muse Code the cheapest credible coding-agent tokens currently on offer, with a data trade attached that many teams cannot make.

The Muse Code complaint that surfaced fastest was not the price but the absence of a ceiling: there is no hard spending cap, only email alerts. Developers on Hacker News described being one mistake away from a large bill, and the workarounds people reached for — virtual cards with low limits, per-key spend limits through OpenRouter — are not Muse Code features. The pricing breakdown works through both tiers and the cost calculator prices a session before you attach a card.

• Windows. There is no native Muse Code build and Meta has said nothing about one. WSL2 runs the Linux binary, but that is a workaround rather than a supported path — see the Windows guide.
• How many skills ship. Meta names three and says "several". The widely-repeated figure of around ten, and the taste skill in particular, remain community claims.
• Whether a plugin system is official. Community reverse-engineering describes hooks and MCP support in the binary. Meta's only related statement is a model-level claim about Muse Spark generalising to new tools and MCP servers, which is not the same thing — the skills page keeps the two apart.
• What "improve their products" covers. The Contributor terms grant training rights, but whether that extends to evaluation, red-teaming or product analytics has not been clarified. The privacy page reads the terms closely.
• Open weights. None released. Zuckerberg said there would be more to share, which is not a commitment.

Benchmarks deserve their own caution. Meta's self-reported Terminal-Bench 2.1 figure of 82.9% sits below Claude Opus 5 and above GPT-5.6 Terra, but it was relayed through press rather than published in a form we could read, and it is not interchangeable with Artificial Analysis's independently measured 80%. A recurring criticism is that the comparison used Terra rather than OpenAI's flagship. The model page keeps the two measurement regimes in separate tables for that reason.

Muse Code is moving fast. When a community claim in the table above becomes documented, when Muse Code prices change, or when Meta answers one of the open questions, this post is updated and the revision date at the top moves with it. Corrections go to hello@muse-code.app and are handled as described in the editorial policy.

For the evergreen version of any topic here, the guide pages are the better destination — the Muse Code guides are re-checked weekly, while this post stays a record of what launch week actually established.