SkillKit is a package manager for AI agent skills. Each coding agent reads skills from its own directory in its own format (SKILL.md under .claude/skills/ for Claude Code, .mdc under .cursor/skills/ for Cursor, Markdown under .github/skills/ for Copilot, and so on across 46 agents), so SkillKit installs from a shared catalog, translates between those formats, and deploys the result to one agent or all of them.
The core loop is four commands: skillkit init detects the installed agent and creates directories, skillkit recommend reads the repository, identifies the stack, and ranks candidate skills with a match percentage, skillkit add installs from a GitHub or GitLab repository, a local path, or a gist, and skillkit sync writes them into the agent's configuration. Installs run a security scan and report how many skills passed. skillkit translate converts a skill to another agent's format, with --all and --dry-run variants.
Four capabilities ship as optional packages so the base install stays small: a terminal interface (@skillkit/tui), a REST server (@skillkit/api), a peer mesh (@skillkit/mesh), and agent messaging (@skillkit/messaging). The CLI prints a one-line hint when a command needs one that is absent.
| Install mode | Packages | Time | Critical/high vulnerabilities |
|---|---|---|---|
| default | 418 | 18s | 0/0 |
| --omit=optional | 118 | 9s | 0/0 |
Beyond installation, skillkit serve runs a REST search endpoint that agents can query at runtime, reachable over MCP through @skillkit/mcp or from Python with SkillKitClient. A programmatic API exposes translateSkill, analyzeProject, RecommendationEngine, MemoryCache, and RelevanceRanker.
Features
- Cross-format translation: one authored skill converts to any supported agent's format, individually or in bulk
- Broad install sources: GitHub and GitLab repositories, local paths, and gists, drawing on a catalog of 400K+ skills across 31 sources
- Stack-aware recommendations: the repository is analyzed and candidate skills ranked by match percentage
- Session memory: skillkit memory compress, search, and export capture what an agent learned during a session and recall it later
- Skill generation: skillkit generate builds a skill from plain English using context from Context7 docs, the codebase, the marketplace, and stored memory
- Runtime discovery: a REST server plus an MCP server let agents fetch skills on demand instead of at install time
- Agent instruction files: skillkit primer --all-agents analyses the codebase and writes CLAUDE.md, .cursorrules, AGENTS.md, and equivalents
- Team manifests: a committable .skills file pins the team's skill set, restored with skillkit manifest install
- Security scanning: skillkit scan checks skills, and installs report a pass count before writing anything
- Peer mesh: encrypted peer-to-peer distribution of agents across machines via skillkit mesh
- Session intelligence: a unified timeline, agent-to-agent handoff, a skill impact graph, and human-readable session summaries
- Browser capture: a Chrome extension saves a webpage as a SKILL.md for later installation
