Browserbase Skills

What It Does

Browserbase Skills is the official skill pack from Browserbase, a cloud browser infrastructure provider. It contains 13 skills that teach AI agents how to browse the web, scrape pages, deploy serverless browser functions, test UIs, capture traces, sync cookies, and research companies. The skills are distributed as a Claude Code plugin via a marketplace.json manifest, and they also follow the agentskills.io convention with per-skill SKILL.md frontmatter. The primary target is developers using Claude Code (or compatible agents) who need browser automation capabilities without writing Playwright or Puppeteer scripts themselves.

The Good

Deep progressive disclosure in every skill. Each skill ships with a concise SKILL.md (the entry point), then expands into EXAMPLES.md and REFERENCE.md files that agents can load on demand. The browser skill, for instance, keeps its 170-line SKILL.md focused on setup and core commands, while REFERENCE.md holds the full API surface. This respects the agent context window while keeping deep docs available.

Real scripts, not just prompts. Nine of 13 skills include .mjs scripts in a scripts/ directory. The browser-trace skill has 8 scripts (start-capture, stop-capture, bisect-cdp, query, snapshot-loop, and more). The company-research skill has 3 scripts (compile_report.mjs, extract_page.mjs, list_urls.mjs) plus a profiles/ directory for persistent user data. These are not stub files: they are 50-200 line Node scripts that do real work.

The ui-test skill is a masterclass in adversarial testing methodology. At 604 lines, it is the longest and most detailed skill in the pack. It defines a three-round planning protocol (functional, adversarial, coverage gaps), structured step markers (STEP_PASS/STEP_FAIL), before/after snapshot comparison patterns, screenshot capture for failures, sub-agent fan-out with explicit step budgets, and an HTML report generator that embeds screenshots as base64. This is more rigorous than most dedicated QA tools.

Thoughtful skill composition. Skills reference each other correctly. The browser-trace skill captures traffic; the browser-to-api skill consumes those captures to generate OpenAPI specs. The cookie-sync skill feeds authenticated sessions into the browser skill for remote testing. The company-research and event-prospecting skills share the same profiles/ directory format. This is a designed ecosystem, not a bag of unrelated tools.

Anti-hallucination rules baked into GTM skills. Both company-research and event-prospecting include explicit rules like "NEVER infer product_description from a site's fonts or framework" and "cap icp_fit_score at 3 when homepage content is not accessible." These are practical guardrails that prevent agents from fabricating research findings.

The Bad

Most skills are API-key gated. Six of 13 skills (fetch, search, functions, cookie-sync, company-research, event-prospecting) require a BROWSERBASE_API_KEY to do anything useful. The remaining skills (browser, browserbase-cli, browser-trace, browser-to-api, safe-browser, ui-test, autobrowse) work in local mode but still benefit from the paid cloud tier. There is no free tier mentioned in the repo, so the on-ramp cost is unclear.

Some skills have hardcoded paths. The event-prospecting skill references /Users/jay/Desktop/ and {SKILL_DIR} with instructions to replace path placeholders. This works for the original author but creates friction for anyone else. The company-research skill has the same pattern with its output directory convention.

No test suite or CI. The repo has no tests directory, no CI configuration, and no validation scripts. For a 3,300-star repo from a funded company, this is a gap. The autobrowse evaluate.mjs script runs real browser sessions for testing, but there are no unit tests for the supporting scripts (compile_report.mjs, extract_page.mjs, etc.).

The autobrowse skill requires an ANTHROPIC_API_KEY. This is the only skill that calls out to a third-party LLM provider directly (it spawns inner Claude agents for iterative improvement). This creates a dependency chain: Browserbase account plus Anthropic account plus Node.js plus Chrome. The skill is powerful but has the highest setup friction.

Some skills have no EXAMPLES.md or REFERENCE.md. The functions skill (119 lines) and browserbase-cli skill (122 lines) are thin, with only a REFERENCE.md and no EXAMPLES.md. Compared to the browser skill (170 lines + EXAMPLES + REFERENCE), they feel underspecified.

Smoke Test Results

Tested on macOS (26.4.1, aarch64) with the browse CLI installed via npx.

Structural validation

$ ls skills/*/SKILL.md | wc -l

13 SKILL.md files found

PASS

$ grep -l "^name:" skills/*/SKILL.md | wc -l

13 of 13 have frontmatter with name field

PASS

$ grep -l "^description:" skills/*/SKILL.md | wc -l

13 of 13 have description field

PASS

$ find . -name "REFERENCE.md" -not -path "./.git/*" | wc -l

10 REFERENCE.md files

PASS

$ find . -name "EXAMPLES.md" -not -path "./.git/*" | wc -l

8 EXAMPLES.md files

PASS

$ find . -name "*.mjs" -not -path "./.git/*" | wc -l

23 scripts (.mjs files)

PASS

$ cat .claude-plugin/marketplace.json | python3 -c "import sys,json; d=json.load(sys.stdin); print(len(d['plugins']))"

5 plugins registered in marketplace

PASS

Pass rate: 7 of 7. All structural checks pass.

browse CLI smoke test (local mode)

$ npx browse@latest --version

browse/0.8.0 darwin-arm64 node-v24.13.1

PASS

$ npx browse@latest open https://example.com --local

{"mode": "managed-local", "title": "Example Domain", "url": "https://example.com/"}

PASS

$ npx browse@latest snapshot

[0-2] RootWebArea: Example Domain

[0-13] heading: Example Domain

[0-18] link: Learn more

PASS

$ npx browse@latest stop

{"stopped": true, "session": "default"}

PASS

Pass rate: 4 of 4. Local browser mode works end to end.

What the runs tell you

The structural validation confirms this is a well-organized multi-skill pack with consistent frontmatter, progressive disclosure files, and real scripts. The browse CLI smoke test proves the core browser automation works in local mode without any API keys. Remote/cloud features could not be tested without a Browserbase account.

Setup Walkthrough

1. Install the browse CLI: npm install -g browse (or run via npx without global install).
2. Verify: browse --version should print browse/0.X.X.
3. For Claude Code integration: /plugin marketplace add browserbase/skills then /plugin install browse@browserbase.
4. For other agents: npx skills add browserbase/skills to install via the agentskills.io convention.
5. For remote Browserbase features: set BROWSERBASE_API_KEY from https://browserbase.com/settings.

The npm install step pulls in several deprecated transitive deps (node-domexception, glob@10.5.0, uuid@10). These are not Browserbase's code but come from the dependency tree. Non-blocking but noisy.

Alternatives

1. Playwright MCP Server -- direct browser control via MCP protocol. No cloud tier, no API key, works with any MCP client. Less turnkey than Browserbase for protected sites.
2. Anthropic's computer-use skills -- screen-level interaction with the OS, not just the browser. Broader scope but less precise for web-specific tasks.
3. Stagehand -- also from Browserbase, but a lower-level SDK. The skills pack wraps Stagehand concepts into agent-consumable instructions. Use Stagehand directly if you want programmatic control without the skill abstraction.