Daedalus

Daedalus extends ToolMaker to automatically discover tasks from Jupyter notebooks and output tools as MCP servers and ToolUniverse wrappers.

What it does

Given a GitHub repository URL, Daedalus:

1. Discovers notebooks and infers task specifications (task.yaml)
2. Builds the tool using ToolMaker's self-correction loop (Docker-isolated)
3. Wraps the result as an MCP server and/or a ToolUniverse tool

uv run python -m toolmaker auto <GITHUB_URL> --name <tool> --output-mcp

Architecture

GitHub Repo
    │
    ▼
DISCOVER ─── analyze_repo() → parse_notebook() → generate_task_spec() → task.yaml
    │
    ▼
BUILD ─────── install (Docker) → create (LLM loop) → validate (tests)
    │
    ▼
WRAP ──────── wrap_as_mcp() → *_mcp.py
              generate_tooluniverse_tool() → *_tu.py + *_config.json

Quick Start

# Clone with submodules
git clone --recursive https://github.com/<you>/daedalus
cd daedalus

# Install dependencies
uv sync

# Configure LLM backend
cat > .env << 'EOF'
TOOLMAKER_LLM_BACKEND=ollama
TOOLMAKER_MODEL=qwen2.5-coder:7b
OLLAMA_BASE_URL=http://localhost:11434
EOF

# Verify setup
uv run python verify_setup.py

# Run full pipeline
uv run python -m toolmaker auto https://github.com/scverse/scanpy \
  --name scanpy_preprocess \
  --output-mcp \
  --output-tooluniverse

Key Concepts

• task.yaml is the canonical intermediate representation — all outputs derive from it
• Tool code executes inside Docker containers, never on the host
• Only open-source LLMs are required (Ollama, vLLM, llama.cpp)
• A single MCP server exposes all tools from a repository (see ADR-0001)

CLI Reference

Command	Description
toolmaker auto <URL>	Full pipeline: discover → build → wrap
toolmaker discover <URL>	Generate task.yaml specs from notebooks
toolmaker wrap <TOOL_DIR>	Wrap a validated tool as MCP server
toolmaker tooluniverse <TOOL_DIR>	Generate ToolUniverse wrapper

See docs/cli.md for full options.

LLM Backends

Backend	Use Case
Ollama	Development, single GPU
vLLM	Production, high throughput
OpenAI-compatible	llama.cpp, LM Studio, LocalAI
LiteLLM	Multi-provider routing

See docs/setup.md for configuration details.

Requirements

• Python 3.12+
• Docker 24.0+
• A running local LLM (Ollama recommended)
• NVIDIA GPU + Container Toolkit (optional, for CUDA tools)

Documentation

• System Overview
• Installation & Setup
• CLI Reference
• Interfaces & Schemas
• Evaluation
• Troubleshooting

Based on

LLM Agents Making Agent Tools Georg Wölflein et al. — ACL 2025 Original repo: KatherLab/ToolMaker