Control your real Chrome browser via HTTP API or MCP server. 185 commands: navigate, click, fill, snapshot, screenshot, assertions, tracing, HAR, video recording, smart caching, token savers, session profiles, humanized input, and more.
Browser Bridge exposes 32 well-organized tools with consistent naming (verb_noun pattern: browser_*) and structured schemas. Most tools have descriptions (194 char avg, within baseline 34-392 range) and input schemas with typed parameters. However, critical gaps exist: (1) Many parameter descriptions are missing or minimal (e.g., browser_find_and_click has NO properties defined in schema); (2) Output schemas are completely undocumented, LLMs cannot predict return structure; (3) Error handling is absent, no guidance on retryability, recovery, or actionable error messages; (4) Several tools lack parameter type constraints (e.g., browser_click_enhanced.button is enum but others like maxClickables lack min/max bounds). Token-saver tools (page_summary, page_diff) are well-described but lack output schema documentation. Overall: solid naming and basic structure, but incomplete parameter documentation and zero output schema coverage prevent higher scores.
Get an accessibility tree snapshot of the page (agent-friendly, labeled interactive elements)
Focus/activate a tab by ID
Click an element by text or selector. Auto-waits + auto-retries (3x with exponential backoff). Scroll-into-view before clicking.
Enhanced click: double-click, right-click, middle-click. Uses real CDP mouse events.
Close a tab by ID
Open a new tab with a URL
Duplicate the current tab
browser_find_and_click has empty inputSchema (no properties defined). LLM cannot determine what parameters to pass.
No output schemas documented for any tool. LLMs cannot predict return structure (e.g., does browser_list_tabs return array of {id, url, title} or {tabId, tabUrl, tabTitle}?). Breaks tool chaining and forces LLMs to guess field names.
Many simple tools (browser_go_back, browser_stop_loading, browser_get_title, browser_get_url, browser_get_meta) lack parameter descriptions. Even empty-param tools should document what they return.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
TOKEN SAVER: Find element by any strategy (text, role, testId, label, selector) + click in one call. Saves 50% vs find + click.
Find element using rich selectors: getByRole, getByText, getByLabel, getByTestId, getByPlaceholder, getByAltText, getByTitle, CSS, XPath, shadow DOM
Get the currently active tab info
Get all attributes of an element
Get the bounding box (position + size) of an element
Get HTML content of an element
Get meta tags from the page (description, og:title, etc.)
Get text content of an element
Get the current page title
Get the current page URL
Go back in browser history
Go forward in browser history
List all open browser tabs
Find element by natural language description (e.g. "blue submit button", "email input next to login"). Uses heuristic scoring — no LLM needed.
Navigate to a URL and wait for page load
TOKEN SAVER: Only return what changed since last page_diff call. First call stores baseline. Saves 95% tokens on repeated inspections.
TOKEN SAVER: Get AI-optimized compact page description (forms, buttons, links, inputs, visible text). ~200 tokens instead of ~2000 for full HTML.
Query a CSS selector (with iframe + shadow DOM support). Returns element info.
Query all matching elements (with iframe + shadow DOM support)
Click using real browser-level mouse events via CDP (for React/Vue/Gmail that ignore synthetic clicks). Auto-waits + auto-retries.
Reload the current page
Get a structured snapshot of the page (clickables, inputs, headings, text). Use lite=true for token-efficient text-only snapshot.
Stop the current page from loading
Wait for the page to finish loading
Wait for the URL to change to match a pattern
Numeric parameters lack min/max bounds. E.g., browser_snapshot.maxClickables, browser_a11y_snapshot.maxDepth, browser_page_summary.maxItems have no constraints. LLMs may pass absurd values (0, -1, 999999) causing timeouts or errors.
Zero error handling guidance. No tool describes what errors are retryable, what user-fixable errors look like, or how to recover. E.g., browser_click fails silently, is it a timeout, element not found, or permission denied? LLM cannot self-correct.