Openbrowser AI

What openbrowser-ai MCP server does

The openbrowser-ai MCP server exposes OpenBrowser’s browser automation framework to MCP-compatible AI assistants. It connects to Chromium-based browsers, including Chrome, Edge, Brave, and Chromium, using Chrome DevTools Protocol. Agents can direct the browser to URLs, interact with page elements, enter text, inspect results, take screenshots, and extract information from web pages.

Its automation model is code-driven rather than a fixed sequence of predefined actions. The LLM can write asynchronous Python for the current task, while OpenBrowser executes that code in a persistent, Jupyter-like namespace. This supports workflows that need several browser steps and state carried between those steps.

The project also documents broader capabilities outside MCP mode, including workflow recording, video recording, a command-line daemon, and direct use of BrowserSession or CodeAgent from Python. The MCP use case is most relevant when an assistant needs to operate a real browser instead of calling a website API.

How it works

OpenBrowser communicates directly with the browser through CDP. A browser session can run headlessly, use a configured viewport, take screenshots, and optionally use a proxy. If a Chromium-based browser is already installed, OpenBrowser detects and uses it. When no supported browser is available and uvx is present, Chromium can be installed automatically on first use; the project also provides an openbrowser-ai install command.

The LLM provider is selected through the surrounding OpenBrowser configuration when using the agent features. Documented provider classes include Google, OpenAI, Anthropic, Groq, AWS Bedrock, Azure OpenAI, OpenRouter, and others. MCP clients named by the project include Claude Desktop, Claude Code, and Codex.

Setup and configuration

Install with pip install openbrowser-ai, uv pip install openbrowser-ai, pipx install openbrowser-ai, or the macOS/Linux and Windows installer scripts. uvx openbrowser-ai --mcp runs MCP mode without a separate package installation. Source installation is also available with the repository’s .[agent] extra.

Python 3.12 or newer is required. Optional extras add agent support, individual LLM providers, all providers, or video recording. Provider credentials are configured with environment variables such as GOOGLE_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY, depending on the selected provider. Browser settings such as headless mode, viewport dimensions, extra Chromium arguments, recording output, and proxy details are configured through BrowserProfile.

Tools and capabilities

The openbrowser-ai MCP server supports these documented browser tasks:

• Navigate to web pages.
• Click page controls and enter text.
• Extract information from pages.
• Capture screenshots for visual analysis.
• Run multi-step automation through asynchronous Python.
• Reuse browser state during a persistent agent session.

The repository’s examples show search, product scraping, flight booking, and form-filling workflows, but successful automation depends on the target site and the instructions given to the agent.

Limitations and notes

The project requires Python 3.12 or newer. It depends on a Chromium-based browser, although installation can be automated when uvx is available. LLM-backed agent operation also requires credentials for the selected provider; the repository does not indicate that one provider is mandatory for every MCP invocation.

The hosted version is described separately from the local package. Its dashboard, saved cookies, scheduled tasks, and related cloud features are presented as hosted functionality or waitlist features, so they should not be assumed to be included in the local MCP server. Browser automation can also be affected by site changes, authentication requirements, and browser configuration.