rosie-skills

A fast, cross-platform package manager for AI agent skills. Use as a CLI or as a typed JS library — same binary either way, via a native build on supported platforms or WebAssembly everywhere else.

CLI

npm install -g rosie-skills
rosie-skills install anthropics/skills

Or one-shot via npx:

npx rosie-skills install anthropics/skills

JavaScript API

The package is ESM-only. Use a namespace import:

import * as rosie from 'rosie-skills';

await rosie.install('anthropics/skills');

const skills = await rosie.list();
//   [{ name: 'pdf', source: 'anthropics/skills', ref: 'main', sha: '...', isReference: false }, ...]

const agents = await rosie.agents();
//   [{ name: 'claude', display: 'Claude Code', detected: true, installPath: '/home/.../skills' }, ...]

All functions return Promises. Failures throw Error with a descriptive message.

Install result

install, installFromLockfile, and update return an InstallResult:

interface InstallResult {
  skills: Array<{
    name: string;
    kind: "skill" | "reference";
    installedAgents: string[];     // e.g. ["claude", "cursor"]
    failedAgents: string[];        // agents the symlink couldn't reach
  }>;
  installedAgents: string[];       // deduped union across all skills
  failedAgents: string[];          // deduped union across all skills
  installedInstruction:            // file rosie wrote the references block to
    | "AGENTS.md"
    | "CLAUDE.md"
    | "GEMINI.md"
    | ".github/copilot-instructions.md"
    | null;                         // null for pure-skill installs
  audit: {                         // structured audit of every change + findings
    schemaVersion: 1;
    command: "install" | "update";
    findings: Array<{              // rosie's own warnings, e.g. tag_rewritten
      severity: "high";
      kind: "tag_rewritten" | string;
      skill: string;
      ref: string;
      oldSha: string;
      newSha: string;
    }>;
    changes: Array<{
      name: string;
      kind: "skill" | "reference";
      source: string;
      ref: string;
      sha: string;
      operation: "install" | "update";
      content: string | null;      // full sanitized body, first-install only
      diff: string | null;         // unified diff, updates only
    }>;
  };
}

const result = await rosie.install('anthropics/skills');
if (result.failedAgents.length > 0) {
  console.warn(`couldn't symlink into: ${result.failedAgents.join(', ')}`);
}

failedAgents is non-fatal: rosie tries every detected agent and reports the misses. The canonical install at .agents/skills/<name>/ still lands and the lockfile entry is still recorded, so a subsequent run after fixing permissions will retry the failed agents.

remove() returns void.

Targeted install

await rosie.install('vercel-labs/agent-skills', {
  agent: ['claude', 'cursor'],
  ref: false,
});

As a reference, not a skill

await rosie.install('vercel/next.js', { ref: true });

Observe progress

The library is silent by default. Pass onLog to receive log events:

await rosie.install('anthropics/skills', {
  onLog: ({ level, message }) => {
    if (level === 'error') console.error(message);
    else if (level === 'info') console.log(message);
    // 'warn' and 'debug' levels also available
  },
});

Remove / update

await rosie.remove('pdf');
await rosie.update();           // update everything in rosie.lock
await rosie.update('pdf');      // update just one

Security defenses

Every install applies content sanitization and structured auditing by default; see docs/security for the full threat model. Each defense can be disabled per call via InstallOptions:

await rosie.install('anthropics/skills', {
  stripComments: false,      // keep markdown comments in reference installs
  stripInvisible: false,     // keep zero-width / bidi / tag-block codepoints
  retagDetect: false,        // skip the tag-rewrite check on `rosie update`
  forceAudit: true,          // print audit on stdout regardless of context
  suppressAudit: true,       // never print audit on stdout (still in result)
});

forceAudit and suppressAudit are mutually exclusive; passing both throws. The audit field on the result is always populated regardless of the emission flags — those flags only control whether the wrapped text is written to stdout by the CLI / bin.ts launcher.

Working directory

Every function accepts a cwd option — equivalent to cd'ing into that directory before running. process.cwd() is restored on exit:

await rosie.install('owner/repo', { cwd: '/path/to/project' });

Mirrors the CLI's --cwd flag.

Skip the lockfile

For ad-hoc installs that shouldn't be recorded in .agents/rosie.lock:

await rosie.install('anthropics/skills', { lockfile: false });

Mirrors the CLI's --no-lockfile flag. Available on install, remove, and update.

Reinstall from lockfile

await rosie.installFromLockfile();

How it works

Three platform binaries (linux-x64, darwin-arm64, freebsd-x64) ship as optional dependencies. On those platforms the CLI execs the native binary directly. On everything else (Windows, linux-arm64, darwin-x64, etc.) the package falls back to an inlined WebAssembly build that does the same work in-process. The JS API always uses the WASM build so you can call it synchronously from Node code without spawning a subprocess.

Supported platforms

| Platform | CLI | JS API | |-----------------|------------|--------| | linux-x64 | native | WASM | | darwin-arm64 | native | WASM | | freebsd-x64 | native | WASM | | Everything else | WASM | WASM |

For native installs on platforms we don't ship a binary for:

• Homebrew (macOS / Linux): brew tap matthewp/rosie && brew install rosie
• Arch Linux: yay -S rosie
• Debian/Ubuntu: see https://github.com/matthewp/rosie
• Source: clone the repo and run cargo build --release

License

BSD-3-Clause