ClaudeMap

Skillbox

A self-hosted, versioned skills library for AI agents: immutable revisions, scoped client keys, an HTTP MCP server plus stdio bridge — and it never executes uploaded skill code.

MCP Serversmcp-serverself-hostedskills-libraryversioningdocker
Visit projectkitze/skillbox

Skillbox is a single-owner, self-hosted application (React, Bun, Hono, PostgreSQL) that turns a pile of agent skills into a governed library. Skills are edited as markdown with immutable revisions, conflict detection and restore; profiles grant skill/bundle access with independent create, update, archive and proposal permissions; client keys are revocable, stored only as hashes, and come with usage reporting and owner-reviewed update proposals. Agents connect over HTTP MCP (with a Node/Bun stdio bridge and checksum-verified CLI downloads), and the base tool set is search_skills, recommend_skills, load_skill, read_skill_file and report_skill_use — write/proposal tools appear according to granted permissions. A bootstrap SKILL.md acts as the agent-side skills-library entry point. Optional task-aware skill recommendations run on your own TypeSafe (Jev) or Vercel AI Gateway key, over a bounded catalog (maximum 200 leaf skills / 120,000 serialized characters) with an uncalibrated 0-4 rubric relevance score; any failure falls back to deterministic PostgreSQL search with an explicit reason. Fetching validates every path, file hash, size, executable flag and package checksum, then writes atomically — the server never runs uploaded skill code. Provider keys are encrypted server-side with AES-256-GCM. Docker-only setup with optional Caddy HTTPS and Umbrel packaging; MIT.

What it does

  • Immutable revisions with conflict detection and restore; bundles expand grants into deduplicated current leaf skills
  • Scoped, revocable client keys stored as hashes only, with per-profile grants and owner-reviewed update proposals
  • HTTP MCP plus a stdio bridge; every fetch validates paths, hashes, sizes, executable flags and package checksums before an atomic write
  • Optional Jev skill recommendations on your own API key, with a hard 200-leaf catalog cap and deterministic PostgreSQL-search fallback

When to use it

Teams and multi-machine individuals who want one governed, versioned source of agent skills — with access control and usage reporting — instead of hand-synced ~/.claude/skills folders on every machine.

Getting started

# Docker Engine/Desktop with Compose v2 + Bash

git clone https://github.com/kitze/skillbox.git
cd skillbox
bash scripts/skillbox.sh setup   # creates .env with random credentials, mode 0600
bash scripts/skillbox.sh start   # then open http://127.0.0.1:4791

# agent side: install bootstrap/SKILL.md, point MCP at /mcp with a client key

Things to know

MIT. Docker-only; a new instance starts completely empty — no seeded skills, accounts or keys. Recommendation relevance is an uncalibrated 0-4 rubric score, not a probability, and catalogs beyond 200 leaves deliberately fall back rather than rank a hidden subset. In Jev mode the task text and authorized skill descriptions go to the selected provider — your key, your bill. Single-owner self-hosted app, not a multi-tenant SaaS; the admin token also derives the key-encryption material, so rotating it invalidates stored provider credentials.

Repository

kitze/skillbox

Charted

Related on the map

  • Blitz Strike

    MCP Server

    A penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.

    • mcp-server
    • security
    • pentest
    • methodology
    View

  • Ripwire

    MCP Server

    Red Hat's "ripgrep for AI context" — a C++23 zero-dependency CLI that also speaks MCP.

    • mcp
    • code-context
    • cli
    • red-hat
    View

  • Deja Vu

    MCP Server

    One shared session memory for 25+ coding agents — Claude Code, Codex, and Cursor finally remember the same project.

    • mcp
    • memory
    • cross-agent
    • context
    View