Agent Browser Protocol
by Independent open-source project (maintainer: theredsix)
Deterministic browser control for AI agents, built into Chromium itself
Agent Browser Protocol is an open-source Chromium fork that gives AI agents deterministic browser control. It embeds an HTTP and MCP server directly in the browser engine and freezes JavaScript and virtual time between actions, so an agent never acts on a page that has already changed underneath it.
Agent Browser Protocol (ABP) is a Chromium fork that embeds an HTTP server inside the browser engine so AI agents can drive a real browser one settled step at a time. Conventional automation stacks fail agents because the model reasons in discrete turns while the page renders continuously: state can change between the moment an agent decides to click and the moment the click lands. ABP closes that gap by turning browsing into a step machine — each request injects native keyboard and mouse input through Chromium's own input path, waits for an engine-defined settled boundary, captures compositor output, returns an event log, then freezes JavaScript and virtual time until the next step. Every action returns before-and-after screenshots, scroll position and viewport, an event log covering navigation, dialogs, file choosers and downloads, cursor position with type information, and execution timing, with optional bounding boxes drawn over clickable, typeable and scrollable elements. It exposes a REST API on localhost:8222, a built-in MCP server that works out of the box with Claude Code, Codex and OpenCode, and a TypeScript SDK; every session records to SQLite with paired screenshots, aimed at building agent training datasets. The project reports 90.5% average on Mind2Web with 85.51% on hard tasks, and claims roughly 2x lower token usage and 2x faster runs than Playwright MCP at about 100ms overhead per action. Written in C++, BSD-3-Clause licensed, 479 GitHub stars and 17 forks; navigation, input, dialogs, downloads and MCP are production-ready while full headless support and action success tracking remain in development.
The team building a browsing agent that keeps failing on dynamic single-page apps, shadow DOM and client-side routing, and that can absorb the cost of running a self-built Chromium fork.
The page is frozen between every agent action, so the class of failure where the DOM moves between the decision and the click simply stops happening.
At a Glance
- Category
- AI Agents & Orchestration
- Pricing
- Free
- Target Market
- CTOs, Enterprise Developers, AI Engineers, QA and Compliance Teams
- Deployment
- Open-source, Self-hosted
- Founded
- 2026
Key Features
- ✓Frozen state between actions
JavaScript execution and virtual time pause after every step, eliminating the race conditions that break agent automation.
- ✓Native input injection
Real keyboard and mouse events go through Chromium's input system rather than synthetic CDP calls, improving fidelity and anti-bot resistance.
- ✓Built-in MCP server
Works out of the box with Claude Code, Codex and OpenCode without a separate bridge or wrapper process.
- ✓Rich per-action response
Every step returns before and after screenshots, an event log, scroll and viewport state, cursor position and timing.
- ✓Element markup overlays
Optional bounding boxes for clickable, typeable and scrollable elements are rendered directly into the returned screenshots.
- ✓SQLite session recording
All actions and screenshots persist locally, giving teams a ready-made dataset for model training or compliance evidence.
- ✓Stateless REST API
Plain HTTP endpoints on localhost:8222 with no WebSocket connection or session handle to keep alive.
Capabilities
Use Cases
- •Agentic web tasks
Lets an LLM browse, fill forms and complete flows without guessing whether the page has finished settling.
- •Compliance and audit testing
Financial and healthcare teams can prove an exact UI flow with a screenshot-backed log of every step taken.
- •Structured data extraction
Scrapes dynamic single-page applications where selector-based tooling breaks on shadow DOM and client-side routing.
- •Replacing brittle Selenium suites
Converts long-lived flaky scripts into a handful of deterministic API calls, per a testimonial carried on the project page.
- •Training-data capture for browsing models
Session recording to SQLite yields paired screenshots and actions for supervised fine-tuning of vision-language agents.
Ideal For
Best For
- ✓Teams building browsing agents that keep failing on dynamic SPAs and shadow DOM
- ✓Compliance and QA teams needing a screenshot-backed record of an exact UI flow
- ✓Researchers training vision-language agents who need paired action and screenshot datasets
- ✓Developers already using Claude Code, Codex or OpenCode who want browser control over MCP
- ✓Anyone replacing flaky Selenium or Playwright scripts inside an agent loop
Not Ideal For
- ✗Production deployments that need vendor support: an independent comparison called maintaining the 51 GB Chromium fork an enormous burden for a single developer handling security patches
- ✗Parallel background automation at scale — full headless support is still on the roadmap and it currently runs headful
- ✗Teams needing auth vaults and encrypted state storage out of the box, where Vercel's agent-browser is the more production-hardened option
- ✗Anyone requiring upstream alignment: the author accepts Google is unlikely to take these changes, so this stays an independent fork in the mould of Brave or Electron
Integrations
Deployment
Market Analysis
Pros
- ✓Solves the agent/browser race condition at the engine level rather than papering over it with waits and retries
- ✓Strong reported benchmark: 90.5% average on Mind2Web with 85.51% on hard tasks
- ✓MCP-native, so it drops into Claude Code, Codex and OpenCode with no glue code
- ✓BSD-3-Clause and fully self-hosted, with no cloud dependency and no per-action billing
Cons
- ✗Maintaining a Chromium fork is the standing risk; an independent comparison called the 51 GB rebase and security-patch burden enormous for a solo maintainer
- ✗Freezing may only defer the problem on SPAs with optimistic UI updates, as one Hacker News commenter argued — the agent still needs async reasoning
- ✗No headless mode yet, which limits running many agents in parallel in the background
- ✗The performance claims (2x fewer tokens, 2x faster than Playwright MCP) are the project's own; a third-party write-up carried no benchmarks at all
- ✗Very young — created January 2026, only 17 forks, and action success tracking is still unfinished
Pricing
Open source (BSD-3-Clause)
$0
- ✓Full Chromium fork
- ✓REST API on localhost:8222
- ✓Built-in MCP server
- ✓TypeScript SDK
- ✓SQLite session recording
Free and BSD-3-Clause licensed with no hosted tier, no seats and no per-action metering; the real cost is operational, because you build, ship and security-patch a forked Chromium binary yourself.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 4 on domains other than github.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Sema4.ai
Enterprise AI agents for knowledge work, running natively in your Snowflake or cloud account
Sapiom
Agent infrastructure that routes, runs and meters AI agents in production
Itential FlowAI
Governed AI agents for network and infrastructure operations, with deterministic execution and full audit trails
OpenAI Presence
Deploy production-grade AI voice and chat agents with enterprise policies, guardrails and evals