An open protocol for building, syncing, and using modular AI ToolKits across agents
The Agent Tool Protocol (ATP) is a modern, cloud-native protocol that enables AI agents to discover, authenticate, and call external tools through developer-defined ToolKits.
Whether you're building AI-native apps, autonomous agents, or agent marketplaces, ATP gives you a lightweight, developer-first way to expose functionality to LLMs—without reinventing wrappers or managing tool routing manually.
ATP was designed to:
- ✅ Make tool usage portable across devices, apps, and agents
- ✅ Remove friction for developers registering and publishing tools
- ✅ Standardize agent ↔ tool communication with built-in auth and lifecycle
- Tool-first – Tools are real Python functions, not wrappers or proxies
- Agent-agnostic – Works with any agent that supports the ATP protocol
- Cloud-sync ready – ToolKits sync across environments and apps
- Developer-first – SDKs provide decorators, metadata, and built-in auth handling
- 📚 Read the Documentation
- 🛠 Install the Python SDK
| Repository | Description |
|---|---|
specification |
Protocol specification and design principles |
python-sdk |
Python SDK for building and publishing ToolKits |
toolkit store |
(Coming soon) ToolKit discovery platform |
examples |
Sample ToolKits for Github, Coingecko, Hubspot, etc. |
We welcome contributions from developers, researchers, and protocol builders.
Got questions or ideas? Join the conversation on GitHub Discussions
ATP is maintained by the ChatATP team and contributors around the world.
We believe AI agents should be as extensible as modern web apps — and ATP is our protocol for making that future real.