OpenBrowser

What does this agent do, and when should you use it?

OpenBrowser (GitHub: billy-enrizky/openbrowser-ai, MIT licensed) is a framework for intelligent browser automation. Its core is a CodeAgent architecture: the LLM writes Python code in a persistent, Jupyter-like namespace that controls the browser through direct CDP (Chrome DevTools Protocol) communication. It ships as a PyPI package (openbrowser-ai) with a CLI daemon, an importable Python library, a built-in MCP server, and plugins/skills for Claude Code, Codex, OpenCode, and OpenClaw. It supports 15 LLM providers (OpenAI, Anthropic, Google, Groq, AWS Bedrock, Azure, Ollama, DeepSeek, etc.) and auto-detects any locally installed Chromium-based browser. The project's benchmarks report 2–6x fewer tokens than browser-use, Playwright MCP, and Chrome DevTools MCP on equal-accuracy tasks. A full-stack FastAPI backend plus Next.js frontend cloud deployment and two reinforcement-learning research studies on browser agents are also included.

OpenBrowser takes a natural-language task or raw code and executes it against a real browser. Via the CLI, uvx openbrowser-ai -p "task" launches an LLM agent, while the -c flag connects to a persistent browser daemon over a Unix socket for direct code execution. In MCP mode (uvx openbrowser-ai --mcp), a single execute_code tool is exposed to clients like Claude Desktop and Claude Code; the LLM writes Python in a persistent namespace calling pre-imported functions (navigate, click, input_text, click_xy, evaluate, download_file, done, etc.). In library mode you instantiate CodeAgent(task=..., llm=ChatGoogle(...)) or drive BrowserSession / BrowserProfile directly (headless, viewport, proxy, video recording dir). Outputs include extracted data, screenshots, exports to Jupyter notebooks or API crawler code, and ffmpeg-recorded session videos. Cookies and login state persist under ~/.config/openbrowser/profiles/; the cloud stack adds real-time VNC streaming, KMS-encrypted saved logins, EventBridge + SQS scheduled workflows, and SES email notifications.

1. A data engineer scraping product listings from e-commerce sites like Walmart, handling pagination and structured extraction
2. An operations user automating repetitive web flows such as flight booking and multi-step form filling
3. A Claude Code user installing the plugin to gain 7 skills: deep research, web scraping, form filling, e2e testing, page analysis, accessibility audit, and file download
4. A QA engineer using the e2e-testing skill to test web apps by simulating user interactions
5. A researcher connecting any AI assistant through the MCP server, which requires no external LLM API keys since the MCP client provides the intelligence
6. A developer recording a browser workflow and exporting it as a Jupyter notebook or API crawler code

How do you install or deploy this agent?

One-line install on macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/billy-enrizky/openbrowser-ai/main/install.sh | sh

Windows PowerShell:

irm https://raw.githubusercontent.com/billy-enrizky/openbrowser-ai/main/install.ps1 | iex

Homebrew:

brew tap billy-enrizky/openbrowser
brew install openbrowser-ai

pip or uv (recommended):

pip install openbrowser-ai

uv pip install openbrowser-ai

No separate browser install is needed: OpenBrowser auto-detects installed Chrome/Edge/Brave/Chromium, and installs Chromium automatically on first run if none is found and uvx is available. Requires Python 3.12+ and at least one LLM API key, e.g.:

export GOOGLE_API_KEY="..."

Install as a Claude Code plugin:

claude plugin marketplace add billy-enrizky/openbrowser-ai
claude plugin install openbrowser@openbrowser-ai

How do you use this agent?

Run a CodeAgent as a Python library:

import asyncio
from openbrowser import CodeAgent, ChatGoogle

async def main():
    agent = CodeAgent(
        task="Go to google.com and search for 'Python tutorials'",
        llm=ChatGoogle(model="gemini-3-flash"),
    )

    result = await agent.run()
    print(f"Result: {result}")

asyncio.run(main())

Execute code directly via the persistent CLI daemon:

uvx openbrowser-ai -c "await navigate('https://example.com')"
uvx openbrowser-ai -c "print(await evaluate('document.title'))"

Variables persist across daemon calls; the daemon auto-shuts down after 10 minutes of inactivity. Manage it with daemon start/stop/status/restart. To connect Claude Code via MCP, add to your project's .mcp.json:

{
  "mcpServers": {
    "openbrowser": {
      "command": "uvx",
      "args": ["openbrowser-ai", "--mcp"]
    }
  }
}

Key environment variables: OPENBROWSER_HEADLESS (default true), OPENBROWSER_ALLOWED_DOMAINS (domain whitelist), OPENBROWSER_USER_DATA_DIR (persistent Chrome profile directory), OPENBROWSER_MAX_OUTPUT (max output characters per execution, default 10000).

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.1.54 v0.1.54: Fix daemon startup hangs with persistent browser profiles 2026-09-21
   Release notes excerpt

   What's New - Fixes daemon startup hangs when a persistent Chrome profile contains a large storage snapshot. - Keeps daemon browser initialization and failed-start cleanup bounded. Details - Managed daemon profiles no longer replay an implicit storage-state snapshot into the same persistent Chrome profile. - Explicit storage-state configuration remains supported. - A stuck cleanup queues the normal forced browser-stop event and cannot hold the startup request indefinitely. - Plugin and MCP metadata are synchronized to version 0.1.54. …

   View on GitHub ↗
2. v0.1.53 v0.1.53: MCP browser ownership and profile isolation 2026-09-21
   Release notes excerpt

   What's New - Prevent concurrent MCP browser instances from interfering with or terminating one another. - Give managed MCP instances isolated browser profiles by default. - Enforce browser ownership during cleanup so only the owning instance can remove its processes. Details - Adds ownership markers and leases for managed browser profiles. - Rejects conflicting ownership of an explicitly shared profile. - Synchronizes the Claude plugin and MCP Registry metadata to version 0.1.53. Full Changelog: https://github.com/billy-enrizky/openbrowser-ai/compare/v0.1.52...v0.1.53

   View on GitHub ↗
3. v0.1.52 v0.1.52: Cap MCP SDK at compatible major 2026-09-20
   Release notes excerpt

   What's New - Fix CI and runtime compatibility by keeping OpenBrowser on the supported MCP SDK major version. - Add a regression test so future dependency updates cannot silently select an incompatible MCP SDK major. Details - MCP dependencies are constrained to >=1.0.0,<2.0.0. - The lockfile is regenerated with the compatible SDK range. - No migration steps are required for existing users. Full Changelog: https://github.com/billy-enrizky/openbrowser-ai/compare/v0.1.51...v0.1.52

   View on GitHub ↗

FAQ