agenthatch 0.8.20

Turn any SKILL.md into a runnable AI Agent

agenthatch

Turn any SKILL.md into a runnable AI Agent.

agenthatch is an Agent Factory that transforms declarative SKILL.md files into fully functional, standalone AI agents. Inspired by the Claude Code SKILL.md specification, agenthatch goes further: it analyzes, reasons about, and generates production-ready agents with tool calling, MCP integration, and multi-turn conversation capabilities.

---

Why agenthatch?

	Claude Code + SKILL.md	agenthatch
Agent format	Inline prompt injection	Standalone runnable agent
Tool calling	Built-in tools only	MCP + custom tools + sandbox
Multi-turn	Single-shot context	Full conversation loop
Deployment	Requires Claude Code	Self-contained Python package
Customization	None	Full harness pipeline
Quality control	Manual	Automated fidelity checks

---

Quick Start

1. Install

pip install agenthatch

2. Initialize

agenthatch init

3. Add a Skill

agenthatch skill add path/to/SKILL.md

4. Hatch an Agent

agenthatch hatch my-skill

This runs the full pipeline:

• Phase 1 — Parse SKILL.md frontmatter and content
• Phase 2 — 6-harness LLM reasoning pipeline (identity, intent, interface, base, assembly, MCP servers)
• Phase 3 — Generate standalone agent code
• Phase 4 — Readiness verification

5. Run the Agent

agenthatch run my-skill

---

How It Works

The Harness Pipeline

agenthatch uses a chain of specialized LLM agents ("harnesses") to analyze and reason about your skill:

SKILL.md
  │
  ├─ Harness A: Identity     → Who is this agent?
  ├─ Harness B: Intent       → What triggers and satisfies it?
  ├─ Harness C: Interface    → What capabilities does it provide/require?
  ├─ Harness F: MCP Servers  → What MCP connections does it need?
  ├─ Harness D: Base         → What runtime environment?
  ├─ Harness E: Assembly     → Cross-validate and produce AHSSPEC
  │
  ▼
agenthatch.yaml (AHSSPEC)
  │
  ▼
Generated Agent (standalone Python package)

Fidelity Protection

Every generated agent includes:

• Fidelity Anchors — SHA256 hashes of constraints extracted from the original SKILL.md
• Fidelity Manifest — Verification file in the agent directory
• Quality Review — Harness E validates intent fidelity, capability coverage, and MCP integrity

Skill Management

# List all skills
agenthatch skill list

# Add a new skill
agenthatch skill add path/to/SKILL.md

# Delete a skill
agenthatch skill delete my-skill

# Search skills
agenthatch search "data analysis"

---

SKILL.md Format

agenthatch follows the Claude Code SKILL.md specification:

---
name: My Skill
description: What this skill does
---

# Skill Instructions

Detailed instructions for the agent...

## Workflow

1. Step one
2. Step two

## MCP Tools

This skill uses mcp__my-server__my-tool for data access.

MCP Support

agenthatch automatically detects MCP server references in your SKILL.md:

• mcp__SERVER__TOOL patterns
• mcporter call Server.Tool syntax
• Frontmatter mcpServers declarations

---

Architecture

agenthatch/
├── src/agenthatch/          # CLI, skill engine, harness, generation
│   ├── cli/                 # Typer CLI commands
│   ├── skill/               # Skill parsing, harness, validation
│   ├── generate/            # Agent code generation + templates
│   ├── agent/               # Runtime, builtins, MCP
│   ├── house/               # Skillhouse index, discovery
│   └── config/              # Configuration management
├── agenthatch-core/         # Universal agent runtime
│   └── src/agenthatch_core/ # LLM client, sandbox, conversation loop
└── tests/                   # Test suite

---

Requirements

• Python 3.11+
• LLM API access (OpenAI, DeepSeek, or custom provider)
• Optional: mcporter for MCP server support (npm install -g mcporter)

---

Contributing

agenthatch is in active development. Contributions are welcome!

# Development setup
git clone https://github.com/agenthatch/agenthatch
cd agenthatch
pip install -e ".[dev]"

# Run tests
pytest

# Quality checks
hatch run quality:check

---

License

MIT — see LICENSE for details.