Browserbase Ships Open-Source Skills So Browser Agents Stop Forgetting Everything
Browserbase open-sources a self-improving browser agent that writes its own playbooks, cutting per-run costs by 45% after just a few iterations
Takeaways
- Browserbase Skills is an open-source Claude Code SDK + skill catalog that gives browser agents persistent memory across sessions, eliminating repeated site re-discovery.
- Autobrowse is a self-improving two-agent loop that iterates on real tasks, studies its own traces, and distills learnings into reusable
SKILL.mdplaybooks. - Benchmarks show a 45% cost reduction on Craigslist scraping ($0.22 → $0.12/run) after just 4 Autobrowse iterations.
- The Browse.sh catalog ships with 200+ pre-built skills covering Amazon, Airbnb, GitHub, federal portals, and more -- installable with one CLI command.
- Skills are plain
SKILL.mdmarkdown files compatible with Claude Code, Cursor, Codex, and Gemini CLI; the repo has 3.6k stars. - The project is fully free and open source; local use requires no Browserbase account, though cloud sessions and hosted browsers need an API key.
Every browser agent you've ever shipped starts from zero. It opens a page, maps the layout, finds the button, and the moment the session closes, it forgets everything. The next run repeats the same exploration at the same token cost. Browserbase just shipped a fix for that.
Browserbase Skills is an open-source collection of agent skills for Claude Code, paired with a self-improving automation system called Autobrowse and a public skill catalog at Browse.sh. The core idea: run a browsing task a handful of times, let the agent study its own failures, and distill what it learned into a plain markdown file. Every future agent reads that file before starting, skipping the discovery phase entirely.
The cost of forgetting
Browserbase benchmarked this on Craigslist. After four Autobrowse iterations, the resulting Browse.sh skill completes the same job for roughly $0.12 per run, a 45% reduction. The savings come from encoding the shortest reliable path the agent found: the undocumented endpoint, the decode tables, the geo-override hack. Every subsequent run reuses that path instead of re-deriving it. At scale, that's the difference between a cost curve that flatlines and one that compounds against you.
Browse.sh launched with 100+ curated browser skills that any agent can install with a single CLI command. Skills are durable, reusable playbooks that capture how to navigate real websites so agents stop re-discovering the same ground on every run.