AgentFactory

The universal AI agent factory — a Python SDK and a self-hosted Studio for building, running, and operating any AI agent.

AgentFactory is a configuration-driven Agent OS: instead of writing a new codebase for every agent you need, you define an agent's identity, tools, and rules as data and the same engine brings it to life — a software engineer, a research analyst, a customer assistant, anything.

• SDK — one pip install gives you LLM failover, persistent memory, native tool calling, streaming, skills, MCP, and human-in-the-loop approvals.
• Studio — a self-hosted multi-user dashboard: sign up, build agents, install tools/skills/MCP servers, run with full observability, and operate from a built-in terminal.

Table of Contents

• ✨ Features
• 🚀 Quick Start
• 🧩 SDK Usage
• 🖥️ CLI Reference
• 🏗️ Architecture
• 🛡️ Security
• 📚 Documentation
• 🤝 Contributing
• 📄 License

✨ Features

Engine

Studio platform

🚀 Quick Start

Option A — Self-host the Studio (full experience)

# Docker (recommended — one container serves API + UI + worker)
docker build -t agentfactory .
docker run -d -p 8000:8000 \
  -e AGENTFACTORY_JWT_SECRET="$(openssl rand -hex 32)" \
  -v agentfactory-data:/data \
  agentfactory

# …or bare metal
pip install 'agentfactory-studio[platform]'
cd web && bun install && bun run build && cd ..
AGENTFACTORY_SPA_DIR="$(pwd)/web/dist" AGENTFACTORY_JWT_SECRET="$(openssl rand -hex 32)" \
  uvicorn agentfactory.app.main:app --host 0.0.0.0 --port 8000

Open http://localhost:8000, sign up, and you're in the Studio: create an agent, add tools/skills/MCP servers and a model connection with your own API key, run a task, and watch it stream — with budget alerts and full observability.

PyPI note: the distribution publishes as agentfactory-studio because the bare agentfactory name is squatted on PyPI (placeholder 0.0.0 release by another author). The Python import package remains agentfactory. Until the first PyPI release, install from source: pip install git+https://github.com/theaaqibjavaid/agent-factory.git.

Option B — SDK only

pip install agentfactory-studio[all]   # all LLM providers
agentfactory init
agentfactory run

See docs/quick-start.md and docs/self-host.md.

🧩 SDK Usage

import asyncio
from agentfactory import AgentFactory

async def main():
    factory = AgentFactory()
    agent = factory.create_agent("Senior")
    result = await agent.run("Implement user authentication")
    print(result)

asyncio.run(main())

Custom tools, skills, MCP servers, and memory integrate through the registry — see the API reference and tool system.

🖥️ CLI Reference

Full reference: docs/cli-reference.md.

🏗️ Architecture

┌────────────────────────────────────────────────────────────────┐
│                        Studio (web/)                           │
│  Dashboard · Agent editor · Tools/Skills/MCP · Terminal ·      │
│  Observability · Approvals · Settings                          │
└──────────────────────────────┬─────────────────────────────────┘
                               │ HTTP / SSE / WebSocket
┌──────────────────────────────▼─────────────────────────────────┐
│                 Platform API (agentfactory/app)                │
│  Auth · Workspaces · Agents · Runs · Proposals · Memories ·    │
│  Tools · Skills · MCP · Models · Marketplace · Terminal ·      │
│  Observability        (SQLite, argon2id, JWT rotation)         │
└──────────────────────────────┬─────────────────────────────────┘
                               │ in-process worker + run broker
┌──────────────────────────────▼─────────────────────────────────┐
│                    Runtime (agentfactory/runtime.py)            │
│  LLM failover · tool sandbox · skills · MCP attach ·           │
│  constitution · guardrails · budget alerts · notifications     │
└────────────────────────────────────────────────────────────────┘

More: docs/architecture.md, docs/design.md.

🛡️ Security

Security is a first-class concern: argon2id password hashing, JWT rotation with revocation, per-IP rate limiting on auth, validated + sandboxed custom tool execution, MCP command/env allowlists, destructive-command guards, path allowlists, and an automated security pipeline (bandit, pip-audit, mypy, coverage gate) in CI.

• Threat model + test plan: docs/security.md
• Reporting vulnerabilities: SECURITY.md

📚 Documentation

Full index: docs/SUMMARY.md

🤝 Contributing

We welcome contributions of all kinds — code, docs, issues, and feedback.

• CONTRIBUTING.md — dev setup, testing rules, PR workflow
• CODE_OF_CONDUCT.md — community standards
• CHANGELOG.md — release history
• SECURITY.md — reporting vulnerabilities

Every PR runs the full gate in CI: tests + coverage ≥ 80%, mypy, ruff, bandit, pip-audit, and the Studio build.

⭐ If AgentFactory helps you build something, star the repo — it tells us the work matters and helps others find it.

📄 License

MIT © 2026 AgentFactory Contributors