chrome.* APIs — to AI agents through direct API calls rather than fragile UI automation.
What is ABP?
ABP inverts the traditional browser automation model. Instead of agents manipulating UI elements through CSS selectors, applications expose a programmatic API that agents call directly:Why ABP?
Modern AI agents can read files, execute commands, and make HTTP requests, but certain capabilities remain inaccessible:| Capability | Why Only Browsers Have It |
|---|---|
| Canvas/WebGL/WebGPU rendering | Requires browser engine and GPU access |
| Authenticated sessions | User’s cookies, localStorage, and OAuth tokens |
| On-device AI (Summarizer, Translator) | Chrome’s built-in AI APIs run locally in the browser |
| Accurate HTML/CSS rendering | Only browsers render CSS correctly |
| PDF generation with CSS | Browser print engines produce vector PDFs with proper fonts |
| Chrome extension APIs | chrome.tabs, chrome.scripting, chrome.bookmarks, etc. only available to extensions |
Project Components
ABP Specification
A complete protocol specification defining how web applications expose capabilities to AI agents — session lifecycle, capability system, message formats, and discovery mechanism.
ABP MCP Bridge
A reference implementation: a generic MCP server that connects any MCP-compatible AI agent to any ABP-compliant web application.
Design Principles
- Browser as Runtime — The browser is a capability provider, not a UI to manipulate
- Explicit Contracts — All operations have defined inputs, outputs, and error conditions
- Progressive Enhancement — Core features work everywhere; advanced features are opt-in
- Transport Agnostic — Works over Puppeteer, postMessage, or WebSocket
- Future-Proof — The capability system accommodates APIs that don’t exist yet
- Bidirectional — Apps can request information from agents, not just respond
- Observable — Long operations report progress; state changes trigger notifications
Relationship to MCP
ABP is complementary to the Model Context Protocol (MCP), not a replacement:| Aspect | MCP | ABP |
|---|---|---|
| Primary use case | Build servers that give AI access to data and tools | Make web apps accessible to AI agents |
| Deployment model | Purpose-built server (local or remote) | Web application running in a browser |
| Typical capabilities | Databases, filesystems, APIs, external services | Browser-exclusive: PDF rendering, canvas, user sessions |
Use Cases
For Web Developers
- Expose your app’s features to AI agents — data visualization, document conversion, image processing, on-device AI, and more
- Make your SaaS app’s features available programmatically without rebuilding an API
- Let agents leverage your app’s authenticated sessions and user data
For AI Agent Developers
- Access browser-exclusive capabilities (canvas/WebGPU rendering, authenticated sessions, on-device AI)
- Use the reference MCP Bridge implementation as a starting point
- Build specialized clients for specific ABP apps
For Researchers & Experimenters
- Explore new ways to leverage browser capabilities in agentic workflows
- Prototype browser-AI interactions without reinventing infrastructure
- Build proof-of-concept integrations between existing web apps and agents
Project Status
Current Version: 0.1.0-alpha The protocol is under active development. We welcome feedback, issue reports, and contributions.Open Source & License
ABP is an open source community project released under the BSD 3-Clause License.This is a freely available specification built for the community. You can use ABP in commercial and non-commercial projects, build implementations, and contribute to the protocol’s evolution.
- 🔧 Building ABP-compatible apps and extensions
- 🤝 Contributing improvements to the specification
- 💬 Participating in design discussions
- 📚 Sharing implementations and use cases
- 🌍 Growing the ABP ecosystem