BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Muse Code 1.4.3 dedicates four of its 20 bundled skills to pulling users off Claude Code and Codex
Four of the 20 built-in skills in Meta's Muse Code 1.4.3 exist to move users off Claude Code or Codex, according to a dev.to teardown. The author never authenticated, so the post covers the import path and runs no coding task.
The Engineer · Build desk
Key figures: Muse Code's importer has five entry points, over unfinished sessions, memory notes, MCP servers, rules and skills. Claude Code's importer is one command, over config and MCP servers. In a dry run, Muse Code's importer reported 8 candidates from 9 skills.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Muse Code's importer | 5 entry points | sessions, memory, MCP, rules, skills | 10 |
| Claude Code's importer | 1 command | covers config and MCP servers | 10 |
| Muse dry-run candidates | 8 of 9 skills | 12 |
What happened
- The import path covers five kinds of state: unfinished sessions, memory notes, MCP server configuration, rules and skills.
- Claude Code ships its own importer, claude import, but its usage line lists only Codex and Gemini as sources and does not accept muse.
- Muse Code launched on 5 August 2026, left beta on 1 September and installs on macOS as one Homebrew cask with a binary called muse.
Why it matters
- exposure A migrated skill that used allowed-tools to restrict itself arrives unrestricted, and a single warning line is the only notice.
- contradiction The post calls the coverage a bid for near-zero switching cost, yet its own dry run shows a field whose permission meaning does not carry across.
- capability Because --dry-run is read-only, a team can run the importer against its own skills directory before authenticating anything.
The import path is a routing table written into the agent's own skill text. According to the dev.to post, the migrate skill sends session transcripts to resume-claude, MCP servers and memory to migrate, and skills to skills import [6]. A separate subcommand, muse skills import --from claude|codex, takes --scope, --dry-run and --force [7]. The entry points are documented inside the skill descriptions, so the model offers them unprompted [5]. Four of the 20 built-in skills is a fifth of the set [19].
The --dry-run flag makes the importer read-only, so it runs without credentials [11]. Without it, the author could not have tested the importer at all. Pointed at nine skills in a real ~/.claude/skills directory, it recognised eight as candidates and printed three diagnostics [12].
Claude Code skills can carry an allowed-tools field in frontmatter, and Muse Code parses it and grants nothing [14]. The warning says the field "is recorded as advisory metadata but is not enforced and grants no tool permissions" [13]. The author calls that "far better than silently honouring a field with different semantics" [18]. We agree. But a single warning line is easy to miss, and it is the only notice a user gets: a skill that used the field to limit what it could touch arrives without the limit [15].
The author reads the coverage as intent, calling it "a product decision to make switching cost approximately zero" [21]. We think that holds for files: the importer moves them at close to zero cost. Permission constraints cost a manual re-check.
Importers are not unique to Meta. The author wrote: "So importers are table stakes in this category, not a Meta tactic." [20] What differs is breadth. Claude Code's importer is one command over config and MCP servers, and Muse Code's spans five entry points [10]. Claude Code's also fails gently. Run against the author's Codex config, it found three MCP servers, listed seven config keys with no mapping and offered to write a reference skill for manual porting [9].
Treat the result as one machine's. The test ran on 2026-10-07 against Muse Code 1.4.3 and Claude Code 2.1.238, installed and inspected without authenticating [1]. The nine-skill result transfers to another directory only if that directory uses similar frontmatter fields and calls similar binaries. The author states the limit directly: "nothing here says a word about whether Muse Code writes better code than Claude Code" [17].
What to watch
- Whether Muse Code starts enforcing or mapping allowed-tools instead of recording it as advisory metadata.
- Whether Claude Code's import usage line adds muse alongside codex and gemini.
- A run of the session, memory and rules import paths on an authenticated install, since the dry run described targets skills.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence68
- Adoption
- Insufficient
- Hype gap0
- Incentives
- Insufficient
- Confidence66
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The author tested on 2026-10-07 with Muse Code 1.4.3 (1.4.3-R5018.1) and Claude Code 2.1.238; both were installed and inspected, but Muse Code was not authenticated, so no task was run.
- [2]
muse skills list --source built-in returns twenty entries.
- [3]
Four of Muse Code's twenty built-in skills exist for one purpose: moving users off Claude Code or Codex.
- [4]
The import path covers unfinished session state, memory notes, MCP server configuration, rules and skills, across five entry points.
- [5]
The import entry points are documented inside the agent's own skill descriptions, so the model offers them unprompted.
- [6]
The migrate skill routes session transcripts to resume-claude, MCP servers and memory to migrate, and skills to skills import.
- [7]
A dedicated subcommand exists: muse skills import --from claude|codex [--scope user] [--dry-run] [--force].
- [8]
Claude Code ships claude import [codex|gemini], also with --dry-run; its usage line lists codex and gemini and does not accept muse.
- [9]
Run against the author's Codex config, Claude Code's importer found three MCP servers, listed seven config keys with no mapping, and offered to write a reference skill documenting the unmapped items for manual porting.
- [10]
Claude Code's importer is one command over config and MCP servers; Muse Code's is five entry points over unfinished sessions, memory notes, MCP servers, rules and skills.
- [11]
--dry-run makes the Muse Code importer read-only, which means it runs without credentials.
- [12]
Run as a dry run against nine skills in the author's ~/.claude/skills, the importer reported candidates:8 installed:0 quarantined:0 skipped:0 failed:0 and three diagnostics.
- [13]
The importer's warning says allowed-tools "is recorded as advisory metadata but is not enforced and grants no tool permissions".
- [14]
Claude Code skills can carry an allowed-tools field in frontmatter; Muse Code parses it, records it, and grants nothing.
- [15]
If a user migrates a skill that leaned on allowed-tools to constrain what it could touch, that constraint is gone and only a warning line says so.
- [16]
Muse Code launched on 5 August 2026 and left beta on 1 September; on macOS it installs as one Homebrew cask with a binary called muse.
- [17]
The post's author wrote that "nothing here says a word about whether Muse Code writes better code than Claude Code".
- [18]
The author calls the allowed-tools warning the right behaviour, "far better than silently honouring a field with different semantics".
- [19]
Four of twenty built-in skills is 20 percent, one fifth of the set.
- [20]
The post's author wrote: "So importers are table stakes in this category, not a Meta tactic."
- [21]
The post's author wrote that the import coverage is "a product decision to make switching cost approximately zero".
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toMuse Code bundles 20 skills, and 4 move you off Claude Code
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Agent credentials and permission scopeFollow
- AI Coding AgentsFollow
- Agent migration toolingFollow