OpenBrowser plugin
A Claude Code plugin with an MCP server and seven skills: the agent drives a real browser by writing Python code
- 279 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent executes arbitrary Python code and controls a browser: it can press buttons, submit forms and download files
- Saved logins give access to your accounts: cookies and localStorage are stored in a profile on disk
- The allowed-domain list is empty by default, so sites are not restricted
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/openbrowserDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent executes arbitrary Python code and controls a browser: it can press buttons, submit forms and download files
- Saved logins give access to your accounts: cookies and localStorage are stored in a profile on disk
- The allowed-domain list is empty by default, so sites are not restricted
This is third-party code. Review the repository files before installing.
What it does
OpenBrowser connects the agent to Chromium directly through the Chrome DevTools protocol. Instead of a set of separate click and type tools, the MCP server offers one `execute_code` tool. The agent writes Python in a persistent namespace, calls functions such as `navigate`, `click`, `input_text`, `evaluate` and `download_file`, and returns only what was asked for. Coordinate actions cover canvases, PDFs and embedded content. The Claude Code plugin installs this server together with seven skills: cited deep research, data extraction, form filling, end-to-end testing, page analysis, a WCAG accessibility audit and file download. The same program works as a command-line daemon: `openbrowser-ai -c` runs code from Bash and keeps variables between calls. You can save logins, record workflows and export them to a Jupyter notebook or API crawler code, and record session video. There is also a Python library, `CodeAgent`, with 15 model providers, and Codex and OpenCode have their own install instructions.
Who it is for. For developers and testers who want the agent as a controllable browser for data extraction, forms, site checks and file downloads.
Good fit when
- You want the agent to collect data from several pages and return a table without flooding the context with page snapshots
- A multi-step login form is filled by the agent after a login was saved in the profile
- A site needs a WCAG check in one request through the accessibility skill
- The agent must download a PDF or CSV using the cookies of the current session
- You need to save tokens on pages and call the browser from Bash with `openbrowser-ai -c`
Not a fit when
- You need a project with corporate support: the repository is run by an independent author
- The agent must not get access to saved logins: profiles keep cookies and localStorage between runs
- You must run on a Python older than 3.12: the package requires that version or newer
- You need cloud schedules and team sharing: the dashboard with saved cookies and schedules exists only in the hosted version, still on a waitlist
Example request
Open the pricing page on a competitor's site, collect plan names and prices into a table and save it as prices.csvLimitations
The package needs Python 3.12 or newer. A separate browser install is not required: any installed Chromium-based browser works (Chrome, Edge, Brave), and if none is found and `uvx` is available, Chromium downloads on first run. You can preinstall it with `openbrowser-ai install`.
MCP mode needs no model keys because the intelligence comes from the client. The Python library needs a key from the chosen provider, for example `GOOGLE_API_KEY`, `OPENAI_API_KEY` or `ANTHROPIC_API_KEY`. The allowed domain list is set with `OPENBROWSER_ALLOWED_DOMAINS`; it is empty by default, meaning no restriction.
Each MCP process gets an isolated profile under `~/.config/openbrowser/profiles` by default. A profile that survives restarts needs the `OPENBROWSER_USER_DATA_DIR` variable. The cloud dashboard with schedules is available only in the hosted version, on a waitlist.
How to disable. In Claude Code disable the plugin through `/plugin` or remove the `openbrowser` entry from `.mcp.json`. Saved profiles and logins live in `~/.config/openbrowser`; delete the folder if you no longer need them.
FAQ
How do I install OpenBrowser in Claude Code?
Run `claude plugin marketplace add billy-enrizky/openbrowser-ai`, then `claude plugin install openbrowser@openbrowser-ai`. The plugin adds the MCP server and seven skills. If you only need the server, put the `uvx` command with the arguments `openbrowser-ai --mcp` into `.mcp.json`.
Why is one `execute_code` tool better than many buttons?
The agent writes short code and gets back only what it asked for rather than the whole page tree after every step. The author claims this uses 3.2 times fewer tokens than Playwright MCP over six tasks. Measurements live in the `benchmarks` folder.
How do I connect OpenBrowser to Codex and OpenCode?
For Codex clone the repository into `~/.codex/openbrowser` and symlink `plugin/skills` to `~/.agents/skills/openbrowser`. For OpenCode the symlinks point to `openbrowser.js` in plugins and to the skills. Then add the MCP server to the project. Or hand the agent a link to `.codex/INSTALL.md` or `.opencode/INSTALL.md`.
Where are saved logins stored and how do I restrict them?
Each MCP process uses an isolated profile in `~/.config/openbrowser/profiles` and seeds cookies and localStorage once from `profiles/default/storage_state.json`. `OPENBROWSER_ALLOWED_DOMAINS` narrows the permitted domains.
Can I call the browser without MCP?
Yes, through the daemon: `uvx openbrowser-ai -c "await navigate('https://example.com')"`. It starts on first call, keeps variables between commands and shuts down on its own after 10 minutes of inactivity. This is also how OpenClaw connects.
Discussion
No questions or reviews yet.
Related
PluginsBrowserbase's official skill set that gives Claude Code web access: browser automation, search, page fetching and UI testing
A separate browser for agents with a skill and MCP: Claude Code and Codex work in your logged-in accounts, and every session is recorded for replay
CLI, skill and Python library for browser control: the agent clicks, fills forms and reads pages over CDP
A browser for agents and people with an MCP server: the agent works in tabs with your logins while you watch and can take over
A skill and CLI from Browser Use: the agent drives your real Chrome over the DevTools Protocol, logged in as you, and writes any missing helper itself
A Claude Code plugin that drives your Chrome over the DevTools Protocol: tabs, forms, page extraction and screenshots with no third-party dependencies
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| OpenBrowser (this entry) | no | 279 | ||
| Playwright MCP | yes | not required | 38,768 | Microsoft's official MCP server where the browser is driven by many separate tools |
| Browser Use | no | 119,424 | a different project's Python library and CLI for agent-driven browsing | |
| agent-browser | yes | not required | 44,790 | a fast Rust CLI built on accessibility snapshots |
README in short
The document lays out the CodeAgent architecture built on CDP, lists install methods and modes for different agents, then covers configuration, model providers, token benchmarks and project structure. The last sections deal with deployment, tests and a study on fine-tuning browser agents.