Loading ToolBench
A Chrome extension that uses your own Chrome browser as an MCP server, enabling AI agents to control and interact with web pages
Universal Browser MCP provides 6 tools with complete input schemas and descriptions. Naming follows verb_noun convention (get_, performance_, chrome_). However, descriptions vary in quality: some are detailed (chrome_read_page, chrome_computer) while others lack actionable context (performance_analyze_insight). Parameter descriptions are present but inconsistent in depth. No output schemas are documented. Error handling guidance is minimal. Tool composition is reasonable but lacks idempotency guarantees and recovery hints.
Use a mouse and keyboard to interact with a web browser, and take screenshots. * Whenever you intend to click on an element like an icon, you should consult a read_page to determine the ref of the element before moving the cursor. * If you tried clicking on a program or link but it failed to load, even after waiting, try screenshot and then adjusting your click location so that the tip of the cursor visually falls on the element that you want to click. * Make sure to click any buttons, links, icons, etc with the cursor tip in the center of the element. Don't click boxes on their edges unless asked.
Get an accessibility tree representation of visible elements on the page. Only returns elements that are visible in the viewport. Optionally filter for only interactive elements. Tip: If the returned elements do not include the specific element you need, use the computer tool's screenshot (action="screenshot") to capture the element's on-screen coordinates, then operate by coordinates.
Get all currently open browser windows and tabs
Provides a lightweight summary of the last recorded trace. For deep insights (CWV, breakdowns), integrate native-side DevTools trace engine.
Starts a performance trace recording on the selected page. Optionally reloads the page and/or auto-stops after a short duration.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
performance_analyze_insight description is vague ('lightweight summary', 'informational only'). LLMs cannot determine when to call this vs other performance tools or what actionable insights it provides.
No error handling guidance. Tools lack recovery hints (e.g., 'If tab not found, call get_windows_and_tabs first'). Agents cannot self-correct on failures.
chrome_computer accepts both 'ref' and 'coordinates' for actions but lacks explicit documentation of mutual exclusivity or precedence rules. Parameter relationships underdocumented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Stops the active performance trace recording on the selected page.
No idempotency guarantees documented. Agents retrying on ambiguous failures risk duplicate side effects (e.g., double-clicking, duplicate interactions).