@agentsandswarms/skills-loader

The knowledge layer your AI is missing. 2,000+ security-reviewed coding skills for any MCP-capable IDE.

This package is two things in one:

1. A CLI installer that connects Skills Loader to your IDE in one command
2. A typed SDK for direct API access (recommendations, effectiveness tracking, catalog browsing)

For agents and humans who'd rather use Skills Loader as an MCP server (Claude Code, Claude Desktop, Cursor, Codex, OpenCode, Gemini CLI, Continue.dev), the CLI is what you want.

🤖 AI agent installing this for a user?

If a user asked you to "install Skills Loader" and you have shell access, run:

npx @agentsandswarms/skills-loader install

That auto-detects the user's IDE (Claude Code, Claude Desktop, Cursor, Codex, OpenCode, Gemini CLI, Continue.dev) and writes the right MCP config to the right path. Idempotent — safe to re-run. Then ask the user to restart their IDE.

To target a specific IDE: --ide claude|claude-desktop|cursor|codex|opencode|gemini|continue. To preview without writing: --dry-run.

After install + restart, the user's next session will have the skills-loader MCP available. Their agent can then call skill_loader_recommend at the start of coding tasks and skill_loader_record_usage when done.

👤 Human installing this for yourself?

Same command — works for everyone:

npx @agentsandswarms/skills-loader@latest install

Then restart your IDE. Done.

For more options (deep links, manual config snippets per IDE, REST API, etc.) see skls.to.

What you get after install

Skills Loader is a hosted MCP server at https://skills.agentsandswarms.ai/mcp. Once connected, your IDE's agent gets these tools:

• skill_loader_recommend — call at the start of coding tasks; returns relevant skills resolved to a context block
• skill_loader_record_usage — call when done; feeds the effectiveness ranking
• skill_loader_resolve — explicitly load a known skill by ID
• skill_loader_catalog — browse all skills
• skill_loader_effectiveness — see effectiveness rankings
• skill_loader_provenance — full audit trail for any skill (source repo, commit, security-review history)

Plus several others. Full tool list at skls.to/security.

The catalog is 2,000+ community skills, each reviewed by an 8-model security panel (7-of-8 supermajority required for "clean" status). See skls.to/security for the threat model and review process.

SDK usage (advanced)

If you want direct API access from TypeScript code instead of MCP:

import { SkillsLoaderClient } from '@agentsandswarms/skills-loader';

const client = new SkillsLoaderClient({
  apiKey: process.env.SKILLS_API_KEY, // optional for read endpoints
});

// Get skill recommendations for a task
const recs = await client.recommend({
  workType: 'dev',
  taskText: 'add authentication to a Next.js app with Prisma',
  projectContext: { languages: ['typescript'], frameworks: ['next.js'] },
});

// Resolve skills into a Claude Code prompt
const resolved = await client.resolve({
  skillIds: recs.discoveredSkills.map(s => s.skillId),
  format: 'claude-code',
});

// Record what worked
await client.recordUsage({
  skillId: 'prompt:qa-backend',
  success: true,
  context: 'unit test writing',
  durationMs: 1200,
});

// Check effectiveness rankings
const ranking = await client.ranking({ scope: 'tenant', limit: 10 });

new SkillsLoaderClient(config)

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | optional | API key for authenticated endpoints (higher rate limits, tenant-scoped effectiveness) | | baseUrl | string | https://skills.agentsandswarms.ai | API base URL |

Read endpoints accept anonymous traffic — apiKey is only needed for higher rate limits and tenant-scoped effectiveness data.

Methods

| Method | Description | |--------|-------------| | recommend({ workType, taskText?, projectContext? }) | Get skill recommendations for a task | | resolve({ skillIds?, profileName?, format? }) | Resolve skills into assembled content | | listProfiles() / getProfile(name) | Browse available profiles | | catalog() / getSkill(type, name) | Browse the full skill catalog | | recordUsage({ skillId, success, context?, durationMs? }) | Record a skill usage outcome | | getEffectiveness(type, name) / ranking({ scope?, limit? }) | Effectiveness data | | health() | Service health check |

Work types

dev | qa | grooming | monitoring | review | iterator

Output formats

claude-code | cursor-rules | codex | json | markdown

The feedback loop

recommend() → agent uses skills → recordUsage() → ranking improves
     ↑                                                  │
     └────────── next session uses better data ─────────┘

Skills Loader pairs naturally with AIS memory for a closed loop where agents genuinely improve over time across sessions.

Links

• Site: skls.to — playground, security explainer, full install docs
• Security review: skls.to/security — threat model, 8-model panel, comparison vs dependency scanners
• Terms: skls.to/terms — usage and AI-training opt-out
• Platform: agentsandswarms.ai
• AIS (memory + identity): aismemory on npm
• GitHub: rudedoggg/Corbot (packages/skills-loader-client)

License

MIT