Autobrowse (Browserbase Skills)

Open-source pattern from Browserbase: a browser agent learns a site over a few iterations, then persists the converged approach as a reusable playbook so the next run skips rediscovery — procedural memory for web tasks.

Current Source Snapshot

As of 2026-08-12, the source authority is browserbase/skills commit 6afe8663693372e59e167dfa5be37932af09ae3d, with 3,687 stars, 235 forks, 16 skill files, one workflow, no detected tests, and GitHub license NOASSERTION. It remains an external procedure source, not a blanket runtime import. Source: commit-pinned repository receipt, 2026-08-12

What it is (compiled)

browserbase/skills is Browserbase's official skill collection for Claude Code and other coding agents, built around the official browse CLI. It is a public, actively-maintained repo. Install with npx skills add browserbase/skills, or add it as a Claude Code plugin marketplace (/plugin marketplace add browserbase/skills). Source: GitHub browserbase/skills, https://github.com/browserbase/skills, 2026-06-30

The repo ships ~11 composable skills. The "learn-once" loop maps to site-debugger, which diagnoses failing automations (bot detection, selectors, timing, auth, CAPTCHAs) and generates a tested site playbook. Supporting skills: cookie-sync carries auth by syncing local Chrome cookies into a Browserbase persistent context; browser runs CDP automation with CAPTCHA solving and residential proxies; fetch / search answer no-session HTTP/web queries; ui-test runs adversarial UI testing off git diffs; browser-trace captures a full CDP trace for debugging; functions deploys serverless browser automation to Browserbase cloud. Source: GitHub browserbase/skills, https://github.com/browserbase/skills, 2026-06-13

Mechanism

1. Agent performs a real task on a live site; fails and retries over ~3–5 loops.
2. site-debugger converges on selectors/timing/auth and writes a playbook that later runs consume before acting.
3. Versus stateless browser loops, repeated workflows amortize discovery cost and latency.

The "learns once, then does the job 10x cheaper forever" framing is announcement/community copy — re-measure cost and latency locally before trusting specific numbers. Source: X/@DeRonin_, 2026-05-08

Relationship to this wiki

• Complements Browser Testing Skills (token-efficient local execution) with knowledge reuse — the persisted playbook artifact.
• cookie-sync is the same idea as Agent Cookie - Session State Sync for the Agent's Second Mac (carry credentials so an agent wakes up authenticated); site-debugger's bot-detection analysis overlaps Web Scraping Stealth (TLS Fingerprinting & Anti-Bot Evasion).
• ui-test fits the Complementary Browser Testing split: agent-driven exploration that later hardens into deterministic gates.
• Aligns with The Brain-Agent Loop and The Resolver Pattern: compile once, reuse everywhere.

Stack fit (Kevin)

• A site playbook is the The Resolver Pattern applied to web tasks — the same "compile once, reuse" thesis the kevin-wiki harness runs on.
• For Dedalus / Agent Machines browser automation, site-debugger + cookie-sync handle the two hardest parts (bot detection, auth) without bespoke glue.
• Browserbase = remote/cloud sessions; default to local Browser Testing Skills for token efficiency and reserve cloud sessions for CAPTCHA / residential-proxy targets.
• cookie-sync overlaps Agent Cookie - Session State Sync for the Agent's Second Mac for headless / second-Mac agents — pick one credential-sync path rather than maintaining two.

Skill Adoption Boundary

Do not import the whole browserbase/skills pack as active runtime state just because the repository is high-signal. The useful parts for Kevin are the pattern and the specific candidate skills: site-debugger for failed-site playbook creation, cookie-sync for authenticated repeat work, safe-browser for domain allowlists, and browser-trace for debugging. Adoption needs a separate audit because the pack is Browserbase-account dependent, Claude-plugin oriented, and can trigger remote browser sessions, proxies, CAPTCHA solving, and cookie movement.

---

Timeline

• 2026-08-12 | Refreshed Browserbase Skills to commit 6afe866, 16 skill files, one workflow, no detected tests, and license NOASSERTION; retained local-browser-first routing and scoped site-playbook candidates. Source: commit-pinned repository receipt
• 2026-06-30 | Deep-reviewed the Browserbase skills bookmark, local screenshot, repo snapshot, and artifact replies. Kept it as a reference/candidate skill pack rather than active local import; the durable rule is learn a site once, save the playbook, and route remote Browserbase use only when local browser automation is not enough. Source: X/@DeRonin_, 2026-05-08; GitHub browserbase/skills, 2026-06-30
• 2026-06-13 | Quality remediation: pinned the pattern to the browserbase/skills repo (~3.5k★, browse CLI, ~11 skills incl. site-debugger playbooks and cookie-sync auth); added stack-fit. Source: https://github.com/browserbase/skills
• 2026-05-27 | @kylejeong says iterative AutoResearch turned browser skills into a /autobrowse command up to 90% faster and cheaper, reinforcing the learn-once-then-amortize pattern. Source: X/@kylejeong, 2026-05-27
• 2026-05-11 | Added to tool registry; ranked under browser automation learning in Skill Resolver. Source: User, 2026-05-11
• 2026-05-08 | Community validation on X - "an agent that learns any website once, then does the job 10x cheaper forever." High engagement (2,061 likes, 4,735 bookmarks) confirms the "learn once, run forever" pattern resonates with developers using browser agents. Source: X/@DeRonin_, 2026-05-08