Lexmount builds browsers and web-access tools for AI agents. Our platform provides Web Search, Web Fetch and managed cloud browsers, with Moli as our browser engine.
Global site · 中国站 · GitHub
Browser Eval compares browsers and web tools through direct tool calls and fixed browser trajectory replay. It favors programmatic scoring where tasks have definite answers, retains evidence for each result, and can turn saved results into local HTML reports readers can inspect, rebuild and share. Optional verification provides a separate check of the results.
Our versioned benchmark datasets are available here:
| Dataset | What it evaluates |
|---|---|
| LexBench-Web-Fetch | Page content, documents, structured records, resource discovery and file delivery |
| LexBench-Web-Search | Scenario search queries with relevance, ranking and explicit source requirements |
| LexBench-WebMainBench | Main-content extraction from frozen HTML, using a reference-corrected upstream edition |
| LexBench-SimpleQA-Verified | Factual retrieval questions and reference answers; composed with scenario queries by Browser Eval |
| LexBench-WCXB | Source-qualified frozen web pages and extraction references |
| LexBench-Headless-Browser | Raw CDP, browser-driver compatibility and web-platform behavior on deterministic fixtures |
| LexBench-Trajectory-Replay | Browser Eval-authored public-web tasks with fixed actions and exact checks |
| LexBench-WebArena-Replay | One official-task catalog, with versioned human-recording and corrected replay definitions |
Each dataset card describes its task scope, use, sources and evaluation limits. Derived datasets retain upstream attribution and document their changes.
The Web Fetch comparison shows how task requirements distinguish successful requests from complete content delivery. The Web Search comparison explains relevance grading, ranking quality, latency and cost with query-level examples. Both articles link the measured populations and evaluation conditions.