Skill Compose is an open-source agent builder and runtime platform for skill-powered agents.
No workflow graphs. No CLI.

Key Capabilities

• 🧩 Skills as first-class artifacts — versioned, reviewable skill packages (contracts, references, rubrics, helpers), not brittle graphs.
• 🧠 "Compose My Agent" workflow — describe what you want; Skill Compose finds/reuses skills, drafts missing ones, and composes an agent.
• 🔌 Tool + MCP wiring — connect tools and MCP servers without hand-writing glue code.
• 🚀 Instant publishing — one click to ship as Web Chat (shareable link) and/or API (integrations-ready endpoint).
• 🛡️ Container-first isolation — run agents in containers (or K8s pods) to keep hosts clean and execution reproducible.
• 🧱 Executors for heavy environments — assign custom Docker images/K8s runtimes per agent (GPU/ML/HPC stacks, custom builds).
• 📦 Skill lifecycle management — GitHub import + one-click updates, multi-format import/export, version history, diff/rollback, and local sync.
• 🔄 Skill evolution from reality — improve skills using feedback + execution traces, with proposed rewrites you can review.
• 🗂️ Skill library organization — categories, pinning, and lightweight discovery to stay sane at 100+ skills.

Examples

Compose Your Agent Describe what you want and let Skill Compose build the agent for you — finding existing skills, drafting missing ones, and wiring everything together.

Evolve Your Agent Improve skills automatically from execution traces and user feedback, review proposed changes, accept the rewrite, and watch your agents and skills get smarter.

Demo Agent: Article to Slides Turn any article or paper into a polished slide deck. The agent reads the content, extracts key points, draft storyboards, and generates presentation-ready slides.

Demo Agent: ChemScout Runs in an isolated execution environment! A chemistry research assistant that searches compound databases, analyzes molecular structures, and summarizes findings into structured reports.

Architecture

Some features shown may still be in development.

Quick Start

Get started with Docker:

git clone https://github.com/MooseGoose0701/skill-compose.git
cd skill-compose/docker
# Default model is Kimi 2.5 (thinking disabled, API key: MOONSHOT_API_KEY), add at least one LLM API key.
# You can also set API KEYs manually in the Web UI "Environment" after launch.
cp .env.example .env
docker compose up -d

Open http://localhost:62600 and click "Compose Your Agent".

Stop services:

cd skill-compose/docker
docker compose down

Build from source (for developers)

cd skill-compose/docker
cp .env.example .env
# Use docker-compose.dev.yaml to build images locally
docker compose -f docker-compose.dev.yaml up -d
# After code changes, redeploy (stop, rebuild, restart):
./redeploy.sh          # all services
./redeploy.sh api      # API only
./redeploy.sh web      # Web only

Cleanup (reset to initial state)

cd skill-compose/docker
# '-v' will remove all data stored in volumes
docker compose down -v

# If you started executor profiles, stop them too
docker compose --profile ml --profile gpu down -v

Resources

• 📚 Full Documentation — Getting started, concepts, how-to guides, and reference
• 🔧 API Reference — Complete REST API endpoints
• 🤖 Models & Providers — Supported LLMs and configuration

Contributing

Found a bug or have a feature idea? Contributions welcome!

License

Apache License 2.0 — see LICENSE for details.