Muse Code adds Wonder support
Updated Meta’s CLI profile with the Wonder model family. Wonder Full requires Power Usage or Maximum Usage. Latest documented CLI: 1.4.6. The index now includes a conservative floor on attributed public commits.
/coding-agentsmuse-code
Meta
Muse Code
Meta’s terminal coding agent for repository changes and CI, with subagents, skills, MCP, and sandboxed command execution.
CLIBetaProprietaryBYOKMCP Yes
Recent product updates
Reviewed
Selected releases and additions, with links to the original announcements.
Attributed output
GitHub public-attribution count, 2026-10-12
568Attributed outputGitHub public-attribution count, 2026-10-12Commits and PRs on public GitHub provably produced by this agent. Exact if it is visible through its own bot or workflow, an estimate if measured through attribution it declares (these can swing 30 to 40 percent between measurement runs at multi-million scale), and a floor either way.Index rank · all agents
#25 of 40 CLI agents
#29Index rank · all agents#25 of 40 CLI agentsRanked by agent-opened PRs in the trailing 4 complete weeks, matching the index. Agents without a tracked PR series follow, ordered by total attributed output. Other filters or sorts change the list position.n/a
Business signal
No sourced business metric
n/aBusiness signalNo sourced business metricSee regional plan pricing
Entry price
Subscription allowance or API tokens
See regional plan pricingEntry priceSubscription allowance or API tokensNot published
GitHub stars
No public repo
Not publishedGitHub starsNo public repoNot published
npm demand/wk
Public npm counter
Not publishednpm demand/wkPublic npm counterWeekly package downloads as an install-demand floor; adjusted when the vendor's main install channel is no longer npm.From the public record
What the public commits reveal
100 commits · 5 repositories
Search window: – · captured Oct 14, 2026 (UTC)
The same commits can credit more than one agent
3%3 / 100 sampled commits
3 of 100 unique sampled commits carry additional agent attribution (Claude). Counting each agent separately can therefore count the same commit more than once. This is shared attribution, not proof of a handoff or which tool did most of the work.
One repository dominates this visible sample
83%83 / 100 sampled commits
karthikkode/kardata accounts for 83 of the 100 distinct attributed commits returned. This makes the sample useful for studying that project’s workflow, but unsuitable as a picture of the agent’s wider adoption.
Specific work you can inspect
DataZooDE/escurel: “demo: one story script for both web stacks (story-s2d.sh folded into story.sh) (#697)”. wicanr2/retro-runtime-study: “DSMI 核心與驅動公開文章（PUB10，#48）”.
This is a sample of publicly attributed work, not total usage or a market-share estimate. Repositories are not customers. Attribution is declared in metadata, not independently verified tool execution; it does not prove exclusive authorship or autonomy.
Method and audit data
Up to 100 returned search results ordered by commit date within the audit window; attribution trailers validated and duplicate SHAs removed across repositories. A recent convenience sample, not a random or exhaustive sample. Commit sample dates use the UTC committer timestamp, matching the search window. Excluded 0 commits with five or more distinct recognized agent/model credits in one message; broad credit lists are not used as evidence of a workflow.
Returned records dated Oct 6, 2026–Oct 11, 2026.
Download the dated audit evidence (JSON)Activity Over Time
Weekly attributed output on public GitHub. Output visible through the agent's own bot or workflow is exact; output measured through declared attribution is an estimate. Window: 2025-02-03 to 2026-10-05.
Market
Developer output
Public GitHub commits attributed to Muse Code
A conservative floor from explicit Muse Code attribution in public commits. The sampled credits identify Meta's coding agent; credits naming only a Muse model are excluded. Default coverage has not been established, and private or unattributed work is invisible.
Last full week (ending 2026-10-11): 417 signed commits (+1713% week over week), search estimate.
Snapshot 2026-10-12, confidence Low, tracked since 2026-06-09. Read the methodology.
- Index rank · all agents
- #29 of 47 · #25 among CLI agents
- Demand share
- No public npm package
- Business signal
- No sourced business metric
- GitHub share
- No public repo
- npm channel
- No public npm package
Pricing And Economics
ExpandHide
Pricing snapshot: 2026-10-12. Check the vendor for current terms.
Subscription access or Meta API billing; Wonder Full requires an upper-tier plan.
- Entry price
- See regional plan pricing
- Top self-serve price
- n/a
- Pricing unit
- Subscription allowance or API tokens
- Token transparency
- Task quota
- Included usage
- Everyday: 10–50 prompts per five hours. High: 5x Everyday; Power: 20x; Maximum: 80x.
- Overage path
- Wait for reset or upgrade. Additional API-key usage is billed separately.
Retail Token Equivalent
Subscriptions publish relative usage, not a guaranteed token allocation. API rates: Wonder $4 input / $20 output; Spark $1.25 / $4.25 per million tokens in Standard mode.
| Plan | Price | Allowance | Unit | Token / Overage Notes |
|---|---|---|---|---|
| Everyday Usage | Shown during onboarding | 10–50 prompts per five hours | Plan usage | |
| High Usage | Shown during onboarding | 5x Everyday | Plan usage | |
| Power Usage | Shown during onboarding | 20x Everyday; includes Wonder Full | Plan usage | |
| Maximum Usage | Shown during onboarding | 80x Everyday; includes Wonder Full | Plan usage |
Summary checked 2026-10-12.
Four subscription tiers or pay-as-you-go with a Meta API key. Plan prices and regional availability are shown during onboarding.
- Subscription credentials apply only to Muse Code; additional API keys are billed separately.
- Wonder Full requires Power Usage or Maximum Usage.
- Subscriptions default to Contributor mode; disable it to keep future content out of model training.
Enterprise And Security
ExpandHide
- SSO / SCIM
- Unknown
- Deployment
- Unknown
- Trains on your code
- Opt-out
- Zero data retention
- Unknown
Compliance
Not published
Admin controls
The CLI executes locally; this does not establish self-hosted inference. Subscriptions default to Contributor mode, which permits model training; it can be disabled. Standard API content is not used for training.
Product
Providers
Meta
Model families
Muse WonderMuse Wonder FullMuse Spark 1.3Muse Spark 1.2
Capabilities
Repo editingTerminal commandsSubagentsSkillsMCPHeadless executionCI integrationOS sandbox
Tradeoffs
- Meta models
- Workflow availability varies by build and platform
- Contributor mode uses shared content for training