Show HN: Alumnium – SOTA Browsing for Claude Code

19 hours ago

Alumnium MCP with Claude Code achieves 98.5% on WebVoyager benchmark, setting a new state of the art.
Alumnium is an open-source project, sharing full results, session transcripts, and code for reproducibility.
Alumnium sits between fully autonomous browser agents and raw browser primitives, offering high-level tools like do(), get(), and check().
Claude Code with Sonnet 4.6 and Alumnium MCP 0.18 with GPT-5 Nano were used in the benchmark.
The benchmark included 610 tasks, with adjustments like restoring 20 tasks and updating date-specific references.
Total cost for Alumnium MCP API calls was approximately $5, less than a cent per task.
Alumnium works primarily from accessibility trees, making vision unnecessary for most tasks.
The experiment shows that custom browser agents and modern browser stacks may not be necessary for great results.
Future plans include submitting results to the Steel browser agent leaderboard and evaluating on other benchmarks.

Hasty Briefsbeta