OpenBrowser - Claude Code Plugin

1# OpenBrowser - Claude Code Plugin
2 
3AI-powered browser automation for Claude Code. Control real web browsers directly from Claude -- navigate websites, fill forms, extract data, inspect accessibility trees, and automate multi-step workflows.
4 
5## Prerequisites
6 
7- **Chrome or Chromium** installed on your system
8- **Python 3.12+** with [uv](https://docs.astral.sh/uv/) package manager
9- **Claude Code** CLI
10 
11## Installation
12 
13### From GitHub marketplace
14 
15```bash
16# Add the OpenBrowser marketplace (one-time)
17claude plugin marketplace add billy-enrizky/openbrowser-ai
18 
19# Install the plugin
20claude plugin install openbrowser@openbrowser-ai
21```
22 
23This installs the MCP server, 5 skills, and auto-enables the plugin. Restart Claude Code to activate.
24 
25### Local development
26 
27```bash
28# Test from a local clone without installing
29claude --plugin-dir /path/to/openbrowser-ai/plugin
30```
31 
32### OpenClaw
33 
34[OpenClaw](https://openclaw.ai) does not natively support MCP servers, but the community
35[openclaw-mcp-adapter](https://github.com/androidStern-personal/openclaw-mcp-adapter) plugin
36bridges MCP servers to OpenClaw agents.
37 
381. Install the MCP adapter plugin (see its README for setup).
39 
402. Add OpenBrowser as an MCP server in `~/.openclaw/openclaw.json`:
41 
42```json
43{
44  "plugins": {
45    "entries": {
46      "mcp-adapter": {
47        "enabled": true,
48        "config": {
49          "servers": [
50            {
51              "name": "openbrowser",
52              "transport": "stdio",
53              "command": "uvx",
54              "args": ["openbrowser-ai[mcp]", "--mcp"]
55            }
56          ]
57        }
58      }
59    }
60  }
61}
62```
63 
64The `execute_code` tool will be registered as a native OpenClaw agent tool.
65 
66For OpenClaw plugin documentation, see [docs.openclaw.ai/tools/plugin](https://docs.openclaw.ai/tools/plugin).
67 
68### Standalone MCP server (without plugin)
69 
70Add to your project's `.mcp.json`:
71 
72```json
73{
74  "mcpServers": {
75    "openbrowser": {
76      "command": "uvx",
77      "args": ["openbrowser-ai[mcp]", "--mcp"]
78    }
79  }
80}
81```
82 
83## Available Tool
84 
85The MCP server exposes a single `execute_code` tool that runs Python code in a persistent namespace with browser automation functions. The LLM writes Python code to navigate, interact, and extract data.
86 
87**Functions** (all async, use `await`):
88 
89| Category | Functions |
90|----------|-----------|
91| **Navigation** | `navigate(url, new_tab)`, `go_back()`, `wait(seconds)` |
92| **Interaction** | `click(index)`, `input_text(index, text, clear)`, `scroll(down, pages, index)`, `send_keys(keys)`, `upload_file(index, path)` |
93| **Dropdowns** | `select_dropdown(index, text)`, `dropdown_options(index)` |
94| **Tabs** | `switch(tab_id)`, `close(tab_id)` |
95| **JavaScript** | `evaluate(code)` -- run JS in page context, returns Python objects |
96| **State** | `browser.get_browser_state_summary()` -- page metadata and interactive elements |
97| **CSS** | `get_selector_from_index(index)` -- CSS selector for an element |
98| **Completion** | `done(text, success)` -- signal task completion |
99 
100**Pre-imported libraries**: `json`, `csv`, `re`, `datetime`, `asyncio`, `Path`, `requests`, `numpy`, `pandas`, `matplotlib`, `BeautifulSoup`
101 
102## Benchmark: Token Efficiency
103 
104<<<<<<< HEAD
105### E2E LLM Benchmark (6 Real-World Tasks, N=5 runs)
106 
107Six browser tasks run through Claude Sonnet 4.6 on AWS Bedrock (Converse API). The LLM autonomously decides which tools to call. All three servers pass **6/6 tasks**. 5 runs per server with 10,000-sample bootstrap CIs. Bedrock API tokens measured from the Converse API `usage` field.
108 
109| MCP Server | Tools | Bedrock API Tokens | Tool Calls (mean) | vs OpenBrowser |
110|------------|------:|-------------------:|-----------:|---------------:|
111| **Playwright MCP** | 22 | 158,787 | 9.4 | **3.2x more tokens** |
112| **Chrome DevTools MCP** (Google) | 26 | 299,486 | 19.4 | **6.0x more tokens** |
113| **OpenBrowser MCP** | 1 | **50,195** | 13.8 | baseline |
114 
115### Cost per Benchmark Run (6 Tasks)
116 
117Based on Bedrock API token usage (input + output tokens at respective rates).
118 
119| Model | Playwright MCP | Chrome DevTools MCP | OpenBrowser MCP |
120|-------|---------------:|--------------------:|----------------:|
121| Claude Sonnet 4.6 ($3/$15 per M) | $0.50 | $0.92 | **$0.18** |
122| Claude Opus 4.6 ($5/$25 per M) | $0.83 | $1.53 | **$0.30** |
123 
124### Per-Task MCP Response Size
125 
126MCP tool response sizes show the architectural difference. Playwright and Chrome DevTools dump full page snapshots; OpenBrowser returns only extracted data.
127 
128| Task | Playwright MCP | Chrome DevTools MCP | OpenBrowser MCP |
129|------|---------------:|--------------------:|----------------:|
130| fact_lookup | 520,742 chars | 509,058 chars | 3,144 chars |
131| form_fill | 4,075 chars | 3,150 chars | 2,305 chars |
132| multi_page_extract | 58,392 chars | 38,880 chars | 294 chars |
133| search_navigate | 519,241 chars | 595,590 chars | 2,848 chars |
134| deep_navigation | 14,875 chars | 195 chars | 113 chars |
135| content_analysis | 485 chars | 501 chars | 499 chars |
136=======
137### E2E LLM Benchmark (6 Real-World Tasks)
138 
139Six browser tasks run through Claude Sonnet 4.6 on AWS Bedrock. The LLM autonomously decides which tools to call. All three servers pass **6/6 tasks**. Token usage measured from actual MCP tool response sizes.
140 
141| MCP Server | Tools | Response Tokens | Tool Calls | vs OpenBrowser |
142|------------|------:|----------------:|-----------:|---------------:|
143| **Playwright MCP** | 22 | 283,853 | 10 | **170x more tokens** |
144| **Chrome DevTools MCP** (Google) | 26 | 301,030 | 21 | **181x more tokens** |
145| **OpenBrowser MCP** | 1 | **1,665** | 20 | baseline |
146 
147### Cost per Benchmark Run (6 Tasks)
148 
149| Model | Playwright MCP | Chrome DevTools MCP | OpenBrowser MCP |
150|-------|---------------:|--------------------:|----------------:|
151| Claude Sonnet ($3/M) | $0.852 | $0.903 | **$0.005** |
152| Claude Opus ($15/M) | $4.258 | $4.515 | **$0.025** |
153 
154### Per-Task Response Size
155 
156| Task | Playwright MCP | Chrome DevTools MCP | OpenBrowser MCP |
157|------|---------------:|--------------------:|----------------:|
158| fact_lookup | 477,003 chars | 509,059 chars | 1,041 chars |
159| form_fill | 4,075 chars | 3,150 chars | 2,410 chars |
160| multi_page_extract | 58,099 chars | 38,593 chars | 513 chars |
161| search_navigate | 518,461 chars | 594,458 chars | 1,996 chars |
162| deep_navigation | 77,292 chars | 58,359 chars | 113 chars |
163| content_analysis | 493 chars | 513 chars | 594 chars |
164>>>>>>> origin/main
165 
166Playwright completes tasks in fewer tool calls (1-2 per task) because it dumps the full a11y snapshot on every navigation. OpenBrowser takes more round-trips but each response is compact -- the code extracts only what's needed.
167 
168[Full comparison with methodology](https://docs.openbrowser.me/comparison)
169 
170## Configuration
171 
172Optional environment variables:
173 
174| Variable | Description |
175|----------|-------------|
176| `OPENBROWSER_HEADLESS` | Set to `true` to run browser without GUI |
177| `OPENBROWSER_ALLOWED_DOMAINS` | Comma-separated domain whitelist |
178 
179Set these in your `.mcp.json`:
180 
181```json
182{
183  "mcpServers": {
184    "openbrowser": {
185      "command": "uvx",
186      "args": ["openbrowser-ai[mcp]", "--mcp"],
187      "env": {
188        "OPENBROWSER_HEADLESS": "true"
189      }
190    }
191  }
192}
193```
194 
195## Skills
196 
197The plugin includes 5 built-in skills that provide guided workflows for common browser automation tasks. Each skill is triggered automatically when the user's request matches its description.
198 
199| Skill | Directory | Description |
200|-------|-----------|-------------|
201| `web-scraping` | `skills/web-scraping/` | Extract structured data from websites, handle pagination, and multi-tab scraping |
202| `form-filling` | `skills/form-filling/` | Fill out web forms, handle login/registration flows, and multi-step wizards |
203| `e2e-testing` | `skills/e2e-testing/` | Test web applications end-to-end by simulating user interactions and verifying outcomes |
204| `page-analysis` | `skills/page-analysis/` | Analyze page content, structure, metadata, and interactive elements |
205| `accessibility-audit` | `skills/accessibility-audit/` | Audit pages for WCAG compliance, heading structure, labels, alt text, ARIA, and landmarks |
206 
207Each skill file (`SKILL.md`) contains YAML frontmatter with trigger conditions and a step-by-step workflow using the `execute_code` tool.
208 
209## Testing and Benchmarks
210 
211```bash
212# E2E test the MCP server against the published PyPI package
213uv run python benchmarks/e2e_published_test.py
214 
215# Run MCP benchmarks (5-step Wikipedia workflow)
216uv run python benchmarks/openbrowser_benchmark.py
217uv run python benchmarks/playwright_benchmark.py
218uv run python benchmarks/cdp_benchmark.py
219```
220 
221## Troubleshooting
222 
223**Browser does not launch**: Ensure Chrome or Chromium is installed and accessible from PATH.
224 
225**MCP server not found**: Verify `uvx` is installed (`pip install uv`) and the MCP server starts (`uvx openbrowser-ai[mcp] --mcp`).
226 
227**Session timeout**: Browser sessions auto-close after 10 minutes of inactivity. Use any tool to keep the session alive.
228 
229## License
230 
231MIT
232