Muse Code, Meta’s terminal coding agent, went live on AIxploria at 04:37 UTC on 18 September 2026. By mid-afternoon the new card showed 6,924 views, 61 upvotes, a “Verified Tool” label, a “Paid” tag and a rank badge reading “#87 in Developer Tools”. Its 4.4 star rating came from twelve votes.
The listing is late to the party. Meta shipped the agent in beta on 5 August and took it out of beta at the start of September, adding three subscription plans, a software development kit and a workflow engine. So the directory card is less an introduction than a snapshot of a product that has already changed twice in six weeks, which makes it a good test of how well a directory keeps up.
We read the card the way this series reads every listing: claim by claim against the vendor’s own documentation, with the rank badge walked against the category it links to and the ranked lists searched for the new entry. Most of it holds. Two numbers do not, and the gap between them changes what the cheapest plan is actually worth.
Table of contents
- What the Muse Code Card on AIxploria Says
- Muse Code in Meta’s Own Words
- Checking the Muse Code Card, Claim by Claim
- Why the Muse Code Subscription Multipliers Matter
- Muse Code Pay-As-You-Go and the Contributor Trade
- Where Muse Code Sits Against Claude Code and Codex
- The Muse Code Badge, the Votes and the Ranked Lists
- What the Muse Code Card Leaves Out for Teams
- Should Your Team Trial Muse Code?
- Muse Code FAQ
- References
What the Muse Code Card on AIxploria Says
The card is a full editorial review rather than a one-line blurb. Its headline claim is that Meta “undercuts Claude Code with parallel sub-agents in the terminal”, and it is filed under three tags: AI Agents, Developer Tools and Latest AI.
The summary line
The blurb reads: “Assign complex tasks on large deposits to multiple persistent agents. They plan changes, write code, and verify the results directly from the terminal.” The word “deposits” is a machine-translation slip for repositories, a reminder that AIxploria publishes in French first and translates.
Five pros and three cons
The pros are parallel sub-agents in isolated Git worktrees, token rates “far below” rival agents, a full event log with crash recovery, easy migration from Claude Code and Codex setups, and a TypeScript SDK. The cons are a young plugin ecosystem, a contributor plan “tied to sharing your code”, and “terminal-only, no dedicated app for now”.
The pricing block
The review lists three subscriptions: Everyday Usage at $5 a month for “roughly 10 to 50 requests every 5 hours”, High Usage at $15 with “three times the allowance”, and Power Usage at $50 with “ten times the allowance”. It quotes Muse Spark at $1.25 per million input tokens and $4.25 per million output tokens, and says the model has been “on version 1.3 since the beta ended”.
The verdict
The card concludes that Muse Code “competes on volume” and that teams running long refactors on a budget “gain a sharp second option to keep next to their usual tools”. That is a fair, measured conclusion. The problem is two of the numbers underneath it.
Muse Code in Meta's Own Words
Before grading the card, it helps to know what Meta itself claims. The launch post on Meta AI Research describes Muse Code as a coding agent that “takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results”.
Persistent background agents
The architectural idea is that Muse Code runs “a simple agent loop plus a set of async background agents” that stay alive for the whole session instead of being spawned per task. Meta says this avoids redundant information gathering and “reduces latency and the need for steering on difficult, multi-step tasks”.
A replayable event log
Every model call, tool run, approval and edit is appended to a local event log, which Meta calls “replay-exact and restart-safe”. After a crash, the agent resumes where it stopped. For long refactors that is a practical feature rather than a talking point, because the most expensive failure in agentic coding is losing an hour of context.
Three bundled skills
The launch shipped three skills: /plan turns a task into an approval-gated plan, /grill “stress-tests that plan until it holds up”, and /goal works toward a stated objective. The card reproduces all three correctly in its command table.
Parallel worktrees
Mark Zuckerberg’s own description, quoted by TechCrunch, is the clearest: “When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees. Your working copy is never touched. In testing we had it build six features for a game simultaneously with no collisions.” Muse Code therefore sells parallelism, not a smarter single loop.
Co-trained with its model
Meta co-trained Muse Spark 1.2 with the harness, using “rejection sampled harness trajectories” and the Muse Code toolset, and tested it on GPU kernel optimisation “over 1,000+ tool calls (up to 24 hours)”. That pairing is why Meta presents the agent and the model as one product.
Checking the Muse Code Card, Claim by Claim
We extracted every checkable statement in the review and compared it with Meta’s launch post, the Muse Code documentation on dev.meta.ai, the Meta Model API pricing page and The New Stack’s report on the general-availability release. Where we could not find a claim on the pages we read, we say so rather than calling it wrong.
| Card claim | What Meta publishes | Verdict |
|---|---|---|
| Parallel sub-agents in isolated Git worktrees | Zuckerberg: sub-agents “working in parallel in isolated worktrees” | Matches |
| Event log with crash recovery | “replay-exact and restart-safe” local event log | Matches |
| $1.25 input, $4.25 output per million tokens | Standard tier: $1.25 input, $4.25 output, $0.15 cached | Matches |
| Contributor plan “over ten times cheaper” | $0.10 input, $0.20 output: 12.5x and 21.25x cheaper | Matches, understated |
| Everyday Usage, 10 to 50 requests every 5 hours | “Send 10-50 prompts every 5 hours” | Matches |
| High Usage: three times the allowance | “5x more usage than the Everyday Usage plan” | Contradicted |
| Power Usage: ten times the allowance | “20x more usage than the Everyday Usage plan” | Contradicted |
| Model “on version 1.3 since the beta ended” | 1.3 is billable, but the docs say “The default model is muse-spark-1.2” | Partly true |
| Official page now lists Windows | “Runs on macOS, Linux, and Windows”, with two features missing on Windows | Matches, with gaps |
| Terminal-only, no dedicated app | Interactive terminal, headless muse exec for CI, and an SDK | True, but narrow |
| Reads CLAUDE.md and Codex skill folders | Not found in the docs pages we read | Not verified |
Seven hold, one understates
The architecture, the standard token rates, the Everyday allowance and the platform list all check out. The contributor discount is described as “over ten times cheaper”, which is true but modest: on output tokens the contributor rate is more than twenty-one times lower, and on cached input it is seventy-five times lower.
Two numbers are contradicted
The High and Power multipliers are where the Muse Code card and Meta part company. Meta’s subscription documentation says High Usage gives “5x more usage than the Everyday Usage plan” and Power Usage gives “20x more usage”. The card says three times and ten times. The New Stack reported the same three-and-ten figures in its general-availability story, so the card appears to have copied press coverage rather than the documentation.
One claim is half right
Muse Spark 1.3 did ship on 2 September and appears on Meta’s pricing page. But the Muse Code installation guide still says “The default model is muse-spark-1.2”. A developer who installs today and changes nothing is on 1.2, not 1.3.
Why the Muse Code Subscription Multipliers Matter
A multiplier sounds like a detail. It is actually the whole value proposition of the upper tiers, because the price of each plan is fixed and only the allowance moves.
Under the card’s figures, the three Muse Code plans scale in straight proportion: $5 buys one unit, $15 buys three, $50 buys ten. Every unit costs $5 whichever plan you are on, which is exactly what The New Stack’s quoted engineer praised as “refreshingly simple”.
Under Meta’s documented figures, the maths is different. High Usage buys five units for $15, or $3 per unit. Power Usage buys twenty units for $50, or $2.50 per unit. The top plan is half the unit price of the bottom one, and that is the arithmetic a budget holder actually needs.
Which figure to trust
Documentation beats a press report, and a press report beats a directory card. Meta’s page is also the one that governs what you are sold, and it says “Available benefits will be shown to you during onboarding before you subscribe”. Read that onboarding screen before paying, because it is the only figure that binds.
Why Meta’s figures look familiar
Five times and twenty times are exactly the multipliers Anthropic and OpenAI use for their $100 and $200 tiers. Meta appears to have mirrored the rivals’ ladder at a quarter of the entry price, which is a sharper competitive move than the card’s flat ladder suggests.
Muse Code Pay-As-You-Go and the Contributor Trade
Subscriptions only cover the Muse Code command-line client. Any other API key under the same Meta Model API account “will be billed through pay-as-you-go”, so teams that script against the model directly will still see token bills.
| Per million tokens | Standard tier | Contributor tier | Standard is dearer by |
|---|---|---|---|
| Input | $1.25 | $0.10 | 12.5x |
| Output | $4.25 | $0.20 | 21.25x |
| Cached input | $0.15 | $0.002 | 75x |
| Requests per minute | 3,000 | 100 | 30x more headroom |
| Tokens per minute | 4,000,000 | 3,000,000 | 1.33x more headroom |
| Your data trains Meta’s models | No | Yes | – |
The hidden cost is throughput
The Muse Code card mentions the training trade but not the rate limit. At 100 requests per minute, the contributor tier has one thirtieth of the standard tier’s request headroom, and limits “apply per team, not per API key”. A single developer will rarely notice. A team running several parallel sub-agents, which is the whole point of the product, may hit that ceiling quickly.
What “sharing your code” means
Meta’s pricing page describes the contributor tier as “heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models”. We covered the wider data question in our report on Meta’s contributor pricing; the short version is that the documentation says little about retention or deletion rights.
What about the subscriptions?
The card’s FAQ says the subscriptions “do not rest on that trade”. The subscription page itself says only that data use “depends on the models you select and is described in the Meta Model API Terms of Service”. That is not a contradiction, but it is less definite than the card, and client or regulated code deserves a read of those terms first.
Two smaller charges
Web search grounding costs $2.50 per 1,000 queries on top of token costs, and Meta states there is “no long-context premium”. The second point is genuinely useful for repository-scale work, where contexts fill up fast.
Where Muse Code Sits Against Claude Code and Codex
The card’s headline frames Muse Code as an undercut of Claude Code. The subscription ladders show how deep the undercut is, and also why a direct comparison is harder than it looks.
| Agent | Entry plan | Upper plans | Published allowance |
|---|---|---|---|
| Muse Code (Meta) | $5 Everyday Usage | $15 and $50 (5x and 20x per Meta’s docs) | 10 to 50 prompts per 5 hours on Everyday |
| Claude Code (Anthropic) | $20 Pro | $100 Max 5x, $200 Max 20x | Stated only as multiples of Pro |
| Codex (OpenAI) | $20 Plus | $100 and $200 Pro tiers (5x and 20x) | 10 to 100 local messages per 5 hours on Plus |
The allowance problem
Headline prices favour Meta by a factor of four at entry. But Anthropic does not publish an absolute prompt count, and OpenAI’s estimate varies with model, context and tools. Meta’s own 10-to-50 range is a fivefold spread. None of these numbers is an apples-to-apples unit of work.
The token-efficiency caveat
The New Stack ran Muse Code and Claude Code through the same three coding tasks and found that Muse Code “consumed substantially more tokens and produced a weaker result on a refactoring task, leaving dead code behind on another”. A cheaper token is not a cheaper outcome if the agent needs more of them, or a second pass to clean up.
Where the undercut is real
For high-volume, low-sensitivity work, such as test generation, documentation passes and mechanical refactors across a large monorepo, Muse Code’s price makes a parallel trial cheap. Meta’s AI chief Alexandr Wang told the Wall Street Journal it “can be an incredibly good option, especially from a cost perspective”, and on price alone that is hard to dispute.
The Muse Code Badge, the Votes and the Ranked Lists
Three cheap checks turn a directory listing from marketing into evidence: the vote count behind the stars, the rank badge against its own category, and whether the ranked lists have picked the tool up yet.
Twelve votes behind 4.4 stars
The page’s structured data gives a rating value of 4.4167 from a rating count of 12. That is more votes than most new cards gather on day one, but it is still twelve people, some of whom may never have installed the agent. Treat it as enthusiasm, not evidence.
A badge that cannot be walked
The badge links to AIxploria’s Developer Tools category, whose page title reads “Page 3 of 22” at twelve tools per page. We read the first three pages, 36 tools, and Muse Code was not among them, which is consistent with a rank of 87. Position 87 would sit on page eight, and every page from four onward returned a Cloudflare challenge.
The controls say the ordinal is real
We opened two cards that do sit on the readable pages. Claude Code is badged “#2 in Developer Tools” and sits second; Gemini CLI is badged “#27” and sits 27th. Both matched exactly, so the badge behaves like a working index here, even though we could not reach slot 87 itself.
Absent from every ranked list
Muse Code scored zero string hits on the Top 100, the Ultimate List and the Free AI page. Its own model, Muse Spark, scored zero on the Top 100 and six on the Ultimate List. For comparison, Claude Code scored nine and six.
A propagation update for the series
GPT-6 Astra, whose card went up on 6 September, read zero on both the Top 100 and the Ultimate List at eight days old. Today it reads 20 and 10. So a new card reached the ranked views somewhere between day eight and day twelve, and Muse Code’s zero at eleven hours is normal rather than damning.
What the Muse Code Card Leaves Out for Teams
A directory review is written for someone deciding whether to try a tool. A team deciding whether to allow it needs a different set of facts, and most of them sit in Meta’s documentation rather than on the card.
Windows is supported, with two gaps
The docs confirm Windows support, but “Voice input is unavailable on Windows” and “Session messaging is unavailable on Windows”. The card’s FAQ promises that sessions “can even message each other between terminals over a local socket”. On a Windows estate, they cannot. Agent-run commands also use PowerShell rather than Bash, which matters for any scripts the agent writes.
Sandboxing is on by default
The overview page says Muse Code runs “with approvals and an OS sandbox on from the first run”, and on Windows the sandbox “may prompt for administrator approval” the first time. That is sound cybersecurity posture, and it is a better argument for the product than anything on the card, but it will also generate help-desk tickets on locked-down machines.
Workflows are not everywhere
The general-availability release added Workflows for staged multi-agent jobs. The docs caution that Workflows “are not available on every build or platform”. Plan any evaluation on the build you will actually deploy.
Subscriptions can be revoked
Meta’s subscription page distinguishes cancellation from revocation: when a subscription is revoked, “you lose access immediately instead of at the end of your billing cycle”, including for payment failure or policy violations. For a tool that sits inside a delivery pipeline, that is a continuity risk worth noting in any AI strategy review.
Headless mode is the real story for CI
The card’s “terminal-only” con undersells the product. muse exec runs a prompt to completion for scripts and CI, and muse serve exposes a versioned session protocol with a TypeScript SDK. For a platform team, that makes Muse Code something to wire into pipelines, not just something a developer opens.
Should Your Team Trial Muse Code?
Put together, the card is accurate on what Muse Code is and slightly stale on what it costs. The two contradicted multipliers understate the value of the upper tiers, and the “version 1.3” line overstates what a default install runs.
A sensible trial shape
Run it on a non-sensitive repository on the standard pay-as-you-go tier for a fortnight, beside whatever agent you already use. Log tokens per completed task, not tokens per request, because The New Stack’s finding suggests the difference between the two is where the saving disappears. Only then compare against the $15 or $50 subscription, using Meta’s 5x and 20x figures.
Where it fits and where it does not
Muse Code fits mechanical, parallelisable work on large codebases, where isolated worktrees and a replayable log earn their keep. It fits less well where code is client-owned or regulated and nobody has read the Model API terms, or on a Windows-only estate that wants session messaging.
Keep the contributor tier out of production
The contributor discount is striking, but it pairs training rights with a 100-request ceiling. Those two facts together point it at prototypes and public code only. If your repositories hold personal data, our data protection team would start there.
For more agent and model coverage, see our AI models, tools and releases hub, our earlier look at Meta’s consumer agent in Rival AI agents, Instinct and Meta’s Muse, both add the ability to make calls, and this morning’s directory check of SurveyMonkey on AIxploria.
Muse Code FAQ
Is Muse Code free?
No. Muse Code is billed per token through the Meta Model API or by subscription from $5 a month. Meta’s pricing page mentions platform free-tier credits for the API, but there is no free Muse Code plan.
How much more does Power Usage give than Everyday Usage?
Meta’s subscription documentation says 20 times the Everyday allowance, and High Usage 5 times. The AIxploria card and some press coverage say 10 and 3 times. Trust the figures shown during onboarding.
Which model does Muse Code use by default?
The installation guide says the default is Muse Spark 1.2. Muse Spark 1.3 is available and billed at the same standard rate, so it can be selected.
Does Muse Code run on Windows?
Yes, according to Meta’s documentation, but voice input and session messaging are unavailable on Windows, and agent commands use PowerShell.
Does Meta train on my code?
On the contributor tier, yes, by design. On the standard tier, Meta says prompts and completions are not used for training. For subscriptions, the documentation refers you to the Model API terms.
References
AIxploria Developer Tools category
Introducing Muse Code and Muse Spark 1.2 (Meta AI Research)
Muse Code documentation (Meta Model API)
Muse Code subscriptions (Meta Model API)
Meta Model API pricing and rate limits
Meta’s Claude Code rival exits beta with three new subscription tiers (The New Stack)
Meta launches Muse Code, an AI agent for large code bases (TechCrunch)
Meta wants to get inside your terminal with its new coding agent (The Register)