Muse Code skills and plugins: status and source for every claim
Read the source column first, because half of what circulates about Muse Code skills has no documentation behind it.
| Feature | Status | Source |
|---|---|---|
/plan — produces a plan you approve before work starts | Documented | Meta launch post |
/grill — stress-tests that plan | Documented | Meta launch post |
/goal — drives work toward a stated goal | Documented | Meta launch post |
| "Several default skills", exact number | Not published | No official count exists |
| Persistent asynchronous Muse Code background agents | Documented | Meta launch post |
| Local append-only event log | Documented | Meta launch post |
| Zero-shot use of new tools, MCP servers, custom skills | Model-level claim by Meta | Muse Spark 1.1 announcement |
Around ten bundled skills including taste | Community-reported, unverified | Posts on X, launch discussion |
| Muse Code plugin system with hooks in the binary | Community-reported, unverified | Community reverse-engineering |
| Muse Code system prompt extracted and reused | Reported by one team | Cline, on X |
Nothing in the community-reported Muse Code rows has been confirmed by Meta or reproduced by us. We update this table when documentation appears; see our methodology for how claims are graded.
Officially documented Muse Code skills: /plan, /grill and /goal
Meta's launch post names three Muse Code skills, and together they describe an approve-then-execute loop.
Meta's launch post for Muse Code and Muse Spark 1.2 says the agent ships with several default skills and names three of them. /plan produces a plan you approve before any work starts. /grill stress-tests that plan. /goal drives the work toward a stated goal.
The sequence is the interesting part. Muse Code skills as Meta documents them cover the weak point of long agent runs: a plan that looked fine in paragraph one and fell apart by step forty. Approving a plan, then attacking it with /grill before execution, is a review step, not a prompt trick.
Meta does not publish how many default skills exist, nor a reference page listing them, so "Muse Code skills" as a documented surface stops at those three names. The first-run walkthrough on our Muse Code install guide shows where these commands sit.
Muse Code background agents and the event log
Two documented Muse Code behaviours matter more than any skill list, because they shape how long runs behave.
Muse Code background agents are described by Meta as persistent and asynchronous: a session stays alive so the agent does not re-gather the same context every turn. That is what makes a multi-hour run plausible, and why Muse Code skills such as /goal have something to drive.
The second is a local append-only event log that makes a session replay-exact and restart-safe after a crash. Both are documented in the same launch post. What that log means for data on your machine is covered on our Muse Code privacy page, and the architecture on what Muse Code is.
What the community has found: Muse Code plugins, hooks and MCP
Every Muse Code skills claim here is community-reported and unconfirmed by Meta, and written that way.
Community reports suggest Muse Code ships with roughly ten bundled skills, not three, among them one called taste, described as a list of design don'ts — what not to produce, rather than a procedure to follow. That description comes from posts on X during launch week and the Hacker News launch discussion. Neither the count nor the taste skill appears in Meta documentation, so treat both as unverified.
Community reports also suggest a plugin system inside the shipped binary, with hooks and MCP server support, found by people inspecting the distributed executable rather than reading docs. Muse Code plugins are not described anywhere official that we could find, and an unreleased code path is not a feature until documented.
The one official MCP signal
One documented claim sits adjacent. Meta's Muse Spark 1.1 announcement claimed zero-shot generalisation to new native tools, MCP servers and custom skills — a model-level claim about Muse Spark, not a statement about a plugin API in the Muse Code binary, but the closest official support the MCP story has.
The Muse Code system prompt extraction episode
One team reported lifting the agent's own instructions onto their harness, and reported large savings on a single task.
The Cline team reported on X that they extracted system-prompt instructions from Muse Code and grafted them onto their own harness, reporting large token and time savings on one task afterwards. We are not reproducing the extracted text here, and no one else has reproduced the claim.
Read the Muse Code claim as one data point, not a benchmark: a single task, self-reported, no controlled comparison. It does suggest that some of what makes Muse Code skills effective lives in prompt design rather than model weights — exactly the sort of claim that needs replication before anyone builds a workflow on it.
Using the model elsewhere: opencode, Cline and OpenRouter
Muse Code skills belong to the harness; the model travels separately, so other tools get one without the other.
Muse Spark 1.2 is available through Meta's model API and is listed on OpenRouter, the usual route into third-party harnesses such as opencode or Cline. What travels is the model: Muse Code skills, the background agents and the event log belong to Meta's own terminal client, so pointing opencode at Muse Spark gives you the model without them.
That distinction matters when comparing agents. Documented skills, hooks and MCP support are a checklist item in the Muse Code vs Claude Code comparison, and the Muse Spark API guide covers access paths. For where the whole field sits, see our best AI coding agents roundup.
Related Muse Code guides
Where to read more about Muse Code skills next: the documented surface, or the wider context.
- What is Muse CodeThe agent, the loop and the model.
- Muse Code install guideFirst run, including the skill commands.
- Muse Spark 1.2Specs and benchmarks for the model.
- Muse Code privacyWhat the event log keeps locally.
- Muse Spark APIUsing the model outside the terminal.
Frequently Asked Questions
What skills does Muse Code ship with?
Meta says Muse Code ships with several default skills and names three: /plan, which produces a plan you approve before work starts, /grill, which stress-tests that plan, and /goal, which drives work toward a stated goal. No total has been published, so any count quoted for Muse Code skills is a community estimate, not documentation.
What is the Muse Code taste skill?
Community reports describe taste as one of the bundled Muse Code skills, holding a list of design don'ts — what not to produce rather than steps to follow. It is not named in Meta's documentation and we have not verified it, so treat it as unconfirmed. If Meta documents the full skill set, this page gets updated.
Does Muse Code support MCP servers?
No MCP support is documented for the Muse Code client itself. The closest official signal is Meta's Muse Spark 1.1 announcement, which claimed zero-shot generalisation to new native tools, MCP servers and custom skills — a model-level claim. Separately, community reports suggest MCP support exists inside the binary, which remains unverified.
Does Muse Code have hooks or a plugin system?
Community reports suggest a Muse Code plugin system with hooks was found inside the shipped binary, and that is the entire basis for the claim. Meta has published nothing about Muse Code plugins, no API is documented, and we have not reproduced the finding. Treat it as a strong hint about direction rather than a feature you can build against today.
What is the Muse Code system prompt?
The Cline team reported extracting system-prompt instructions from Muse Code and applying them to their own harness, reporting large token and time savings on one task. We do not reproduce the extracted text. It is a single self-reported result without a controlled comparison, so it is best read as a hint that prompt design carries real weight here.
How do Muse Code background agents work?
Meta describes them as persistent asynchronous agents: the session stays alive so context is not re-gathered every turn, which is what makes long unattended runs practical. Paired with the local append-only event log, a crashed session can restart and replay exactly. Both behaviours are documented in Meta's launch post, not community-reported.
Can I use Muse Code skills in opencode or Cline?
No. Muse Spark 1.2 is reachable through Meta's model API and OpenRouter, so other harnesses can use the model, but Muse Code skills, background agents and the event log belong to Meta's own terminal client. Pointing another tool at the model gives you the model alone, with that tool's own skill or plugin system around it.
Sources
- Meta AI Research — Introducing Muse Code and Muse Spark 1.2research.meta.ai
- Meta AI — Muse Spark 1.1 announcementai.meta.com
- Meta for Developers — Muse Spark model pagedeveloper.meta.com
- Hacker News — Muse Code launch discussion (278 points, 177 comments)news.ycombinator.com
- OpenRouter — meta/muse-spark-1.2 listingopenrouter.ai