Self-hostedMCP nativeMIT licensed
A skills library for your AI agents
Write a skill once. Version it like code. Serve it to every agent you run — over MCP, CLI, or HTTP. On your own box, under your own keys.
Write a skill once. Version it like code. Serve it to every agent you run — over MCP, CLI, or HTTP. On your own box, under your own keys.
A markdown playbook plus any files it needs — the stuff that teaches an agent a workflow it doesn't know. Skillbox stores them, versions them, and hands them to agents on demand, pinned to exact revisions.
Claude, Codex, Cursor — anything that speaks MCP. Skillbox ships an HTTP MCP server, a stdio bridge, and a CLI with checksum-verified downloads. Agents search the catalog, load pinned revisions, and report what they actually used.
Every publish is an immutable revision. Conflict detection stops silent overwrites, restore brings anything back, and agents can propose changes that you review before they land. Nothing shifts under your agents' feet.
Docker-only setup that starts completely empty — no seeded accounts, no phoning home. Client keys are shown once and stored as hashes. Provider keys are encrypted with AES-256-GCM. And Skillbox never executes uploaded skill code.
Small surface, sharp tools. No accounts, no cloud, no telemetry.
Skill and file editor with frontmatter, plus native folder import and export.
Pinned versions, conflict detection, and one-click restore of anything.
Per-client keys shown once, stored hashed. Profiles scope exactly what each agent can see and do.
HTTP MCP endpoint plus a Node/Bun stdio bridge for clients that need it.
Task-aware skill ranking with your own AI key. Falls back to plain search — and says so.
Protected PostgreSQL and config backups with explicit, deliberate restore tooling.
Docker Compose v2 and Bash. That's the whole requirements list — no host Node, no host Bun. Setup generates unique random credentials into a private .env and refuses to overwrite it.