#########################################################################
# __ #
# _ _ _____ _ _____ __ / \ Helix #
# | | | || ____| | |_ _\ \/ / / /\ \ v1.0.0 #
# | |_| || _| | | | | \ / | | | | -------------- #
# | _ || |___| |___ | | / \ \ \/ / [STATUS: ACTIVE] #
# |_| |_||_____|_____||___/_/\_\ \ / #
# / \ #
# Skills that evolve on demand. / /\ \ #
# | | | | #
# \ \/ / #
# \ / #
# \/ #
#########################################################################
Skills that evolve on demand.
Helix is a self-extending AI agent. Give it a task, and it will check if it already has the right skill — if not, it writes the code, builds a Docker container, deploys it as a microservice, and calls it to return your result. Each skill persists and is reusable for future tasks.
Helix supports multiple LLM providers. We ran the same three prompts through Claude and GPT OSS to compare the skills each model generated.
| Claude | GPT OSS |
|---|---|
 |
 |
| Claude | GPT OSS |
|---|---|
 |
 |
| Claude | GPT OSS |
|---|---|
 |
 |
helix/
├── main.py # CLI entry point
├── config.py # Centralized settings (env vars, constants)
├── orchestrator/
│ ├── agent.py # Claude API loop + tool definitions
│ └── registry.py # In-memory skill registry
├── skill_factory/
│ ├── factory.py # Code generation + Docker build/run
│ ├── port_manager.py # Dynamic port allocation
│ └── templates/
│ └── fastapi_skill/
│ ├── main.py.j2 # Jinja2 template for skill code
│ └── Dockerfile.j2
├── models/
│ └── skill.py # Skill and SkillSpec Pydantic models
├── integrations/
│ └── telegram_bot.py # Telegram bot interface
└── e2e/ # End-to-end tests
- Python 3.13+
- Docker (running)
- uv package manager
- Anthropic API key
# Clone and install dependencies
git clone https://github.com/danielgraviet/helix.git
cd helix
uv sync
# Create Docker network
docker network create agent-net
# Pre-pull base image (speeds up first skill build)
docker pull python:3.12-slim
# Set environment variables
cp .env.example .env
# Add your ANTHROPIC_API_KEY to .envuv run python -m main# Add TELEGRAM_BOT_TOKEN to your .env file
uv run python -m integrations.telegram_bot| Command | Action |
|---|---|
| Type any task | Helix creates/calls skills to solve it |
quit / exit |
Shut down and clean up containers |
| Command | Action |
|---|---|
/start |
Introduction |
/skills |
List active skills |
/clear |
Remove all skills and containers |
| Any message | Helix processes the task |
All settings live in config.py:
| Setting | Default | Description |
|---|---|---|
MODEL |
claude-sonnet-4-5-20250929 |
Claude model for the agent loop |
MAX_TOKENS |
4096 |
Max response tokens |
DOCKER_NETWORK |
agent-net |
Docker bridge network for skills |
SKILL_BASE_IMAGE |
python:3.12-slim |
Base image for skill containers |
PORT_RANGE_START |
9001 |
Start of dynamic port range |
PORT_RANGE_END |
9100 |
End of dynamic port range |
MAX_BUILD_RETRIES |
3 |
Retry attempts for failed builds |
CONTAINER_TIMEOUT |
60 |
Build/run timeout (seconds) |
SKILL_STARTUP_TIMEOUT |
15 |
Max wait for /health (seconds) |
- Agent: Claude API (Anthropic) with native tool use
- Skill Runtime: FastAPI + uvicorn in Docker containers
- Templating: Jinja2
- Container Management: docker-py SDK
- CLI Output: Rich
- Telegram: python-telegram-bot
- Models: Pydantic