Browserbase Skills - Setup Guide¶

Source: browserbase/skills via skills.sh - 40.1K combined installs across 20 indexed listings; first seen 2026-10-31 (morning sweep)

GitHub: browserbase/skills - 3,747 stars, no LICENSE file, branch main, last pushed 2026-10-27; 18 SKILL.md files, one per directory under skills/, ten of which also carry a scripts/ folder of Node standard-library helpers

Category: Browser automation for agents: Browserbase platform, sessions, UI testing and grading

Quality Tier: 🟡 Yellow - one sampled skill fails the Gen Agent Trust Hub auditor, three carry Snyk warns, one carries a Socket warn, and the repository has no LICENSE file at its root

Browserbase ships this as its first-party skill set, and the shape is unusual for a vendor repository: it is not a wrapper around one SDK call but a working browser harness that an agent can drive end to end. The eighteen skills split into four groups. The core loop is browser (natural-language control of a browser through the browse CLI), fetch and search (no session at all), and browser-trace (a read-only CDP trace attached to whatever else is running). Around that sit the quality tools: ui-test for adversarial UI testing, agent-experience for grading how agent-friendly a product or docs site is, and optimize-agent-prompt for tuning a Browserbase Agent system prompt. Then the deployment and interop layer: functions, safe-browser, cookie-sync, webmcp-gen, add-webmcp, browser-to-api, browser-use-to-stagehand and autobrowse. The last group is not browser plumbing at all: company-research, competitor-analysis and event-prospecting are go-to-market research skills that use the Browserbase Search API as their data source.

The dependency to understand before installing is the browse CLI. Most skills call it, and their own compatibility lines ask for npm install -g browse plus BROWSERBASE_API_KEY for remote sessions, or a local Chrome or Chromium for local ones. The two sessionless skills, fetch and search, need the API key and nothing else. That means the suite is closer to a toolchain install than a markdown drop-in: useful immediately for a profile that already drives a browser, and a second dependency to manage for one that does not.

Installation¶

The README's primary route is the skills CLI:

npx skills add browserbase/skills

Claude Code users get a plugin marketplace route as well, which is also documented in the README and installs the browse plugin rather than loose skill folders:

/plugin marketplace add browserbase/skills
/plugin install browse@browserbase

The README then points at the CLI the skills actually call. Each skill's compatibility line names it, and remote sessions need a key:

npm install -g browse
export BROWSERBASE_API_KEY="your-key"

Neither route knows about Hermes, so the durable install for this profile is a directory copy. Clone shallow and copy the skill directories, not the files, so each SKILL.md keeps its scripts/ neighbours:

git clone --depth 1 https://github.com/browserbase/skills.git /tmp/browserbase-skills
cp -r /tmp/browserbase-skills/skills/. ~/.hermes/profiles/corpusiq/skills/

That copies all 18 directories in the tree. A narrower install is reasonable here, because the research skills pull in report templates and the browser skills do not: take browser, fetch, search and browser-trace for the core loop, add ui-test and safe-browser for testing work, and add company-research, competitor-analysis and event-prospecting only if the profile does prospecting.

What It Provides¶

Skill	Installs	Use For
browser	5,909	Driving a browser in natural language through the `browse` CLI: navigation, clicks, form fills, screenshots and data extraction, on local Chrome or a remote Browserbase session with Identity, verified browsers, CAPTCHA solving and residential proxies.
fetch	3,058	Pulling HTML or JSON from static pages with no browser session: status codes, headers, redirect control and page source for light scraping.
browser-trace	3,039	Attaching a read-only CDP client to a session that is already being driven, recording the DevTools firehose, screenshots and DOM dumps, then slicing them into per-page buckets for search and debugging.
ui-test	3,032	Adversarial UI testing over the browse CLI: diff-scoped runs for a pull request or full exploratory passes, checking functional correctness, accessibility, responsive layout and UX heuristics.
functions	2,854	Deploying browser automation as Browserbase cloud functions: scheduled runs, cron jobs and webhook endpoints built with the `browse` CLI.
search	2,829	Web search without a browser session, returning structured titles, URLs, authors and dates.
autobrowse	2,767	The self-improving loop: run a browsing task, read the trace, rewrite the navigation strategy, and repeat until the run passes reliably, including parallel runs across tasks.
cookie-sync	2,665	Exporting cookies from local Chrome into a Browserbase persistent context so the CLI can reach sites you are already signed into.
browserbase-cli	2,489	Indexed on skills.sh with no matching directory in the clone, so there is no `SKILL.md` to route to. Treat it as a pointer to the `browse` CLI that the other skills document, not as a separate install.
company-research	2,365	Finding companies that match an ICP with the Search API, deep-researching each one, and scoring fit into a report plus CSV, with quick, deep and deeper depth modes.
event-prospecting	2,092	Extracting speakers from a conference page, filtering their companies against the ICP, and deep-researching the best-fit people into a person-first report with public links and a DM opener.
browser-to-api	1,957	Turning a site's observable HTTP traffic into a best-effort OpenAPI 3.1 document plus a coverage report, by replaying a `browser-trace` capture.
safe-browser	1,741	Generating a local Claude Agent SDK browser agent whose only browser capability is a CDP-gated `safe_browser` tool with domain allowlist enforcement through Fetch interception.
agent-experience	876	Grading how agent-friendly a product, SDK or docs site is: subagents get a tiny prompt, their traces are captured, and setup friction, speed, error recovery and doc quality are scored A to F.
competitor-analysis	675	Discovering competitors through the Search API, researching each across marketing, signal, benchmark and strategic-diff lanes, and compiling an HTML report with a feature and pricing matrix.
webmcp-gen	574	Authoring, compiling and validating site-specific WebMCP init scripts from a target URL, producing a `webmcp.init.js` artifact and testing registration with Stagehand.
browser-use-to-stagehand	503	Porting browser-use Python scripts to Stagehand v3 on Browserbase, mapping `act`, `extract`, `observe` and `agent` primitives and choosing a determinism level per step.
what-antibot	275	Indexed on skills.sh with no directory in the clone. There is no `SKILL.md` behind it here, so there is nothing to install and nothing to route on.
optimize-agent-prompt	229	Tuning a Browserbase Agent `systemPrompt` in an outer loop: fixed task, collected Agent messages and session logs, a score, one revised heuristic, and confirmation runs.
add-webmcp	198	Adding first-party WebMCP tools to an existing web app by finding safe capabilities across routes, forms, server actions and schemas, then validating discovery with Stagehand.

All 20 indexed listings clear the threshold and appear above. Two of them, browserbase-cli and what-antibot, have no directory in the tree, so the install copies 18.

Why This Matters for Hermes Agents¶

Most agent browser work fails on the last mile rather than the first. The agent can navigate, but it cannot show what happened when navigation broke, cannot prove that a form submission was the right one, and cannot repeat a run that worked yesterday. This suite is built around that gap: browser-trace turns a session into a searchable record, autobrowse closes the loop by rewriting the navigation strategy from that record, and ui-test treats a diff as the scope of the test rather than the whole application.

The second thing it changes is reach. browser is documented as supporting remote sessions with CAPTCHA solving, residential proxies and verified browsers, which is the difference between an agent that can read public pages and an agent that can complete a task on a site that resists automation. cookie-sync extends that to sites the operator is already signed into, and safe-browser pulls in the other direction, generating an agent whose only browser capability is allowlisted by domain, with Fetch interception failing anything off-list.

The third is that half the repository is about grading rather than browsing. agent-experience drops subagents at a product or docs site with a tiny prompt and scores what happened, and optimize-agent-prompt does the same for a Browserbase Agent system prompt. For a profile that publishes anything agent-facing, those two are a feedback mechanism that does not otherwise exist.

Finally, the research skills turn the browser harness into a lead pipeline. company-research, competitor-analysis and event-prospecting all use the Search API for discovery and deep research for the per-entity work, and all three emit reports rather than raw output, which is what makes the results reviewable rather than just voluminous.

Usage¶

You say	What happens
"Go to Hacker News, pull the top post comments and summarise them"	`browser` drives the page through the `browse` CLI, extracts the comments and returns the summary rather than the raw DOM.
"QA test localhost:3000 and fix what breaks"	`ui-test` walks the app over a local browser, checking function, accessibility, responsive layout and UX, then reports bugs it can reproduce.
"Why did last night's automation run fail"	`browser-trace` reads the recorded CDP stream for that run, bisects it per page, and answers from network, console and DOM events instead of guessing.
"Sync my Chrome cookies so the agent can browse a signed-in site"	`cookie-sync` exports cookies into a Browserbase persistent context, and later `browse` calls reuse that context.
"Turn this dashboard's XHR traffic into an OpenAPI spec"	`browser-to-api` pairs request and response events from a trace, templatizes the observed URLs and emits OpenAPI 3.1 with a coverage report.
"Find fifty companies that match this ICP and score them"	`company-research` discovers candidates through the Search API, deep-researches each, and writes a scored report plus a CSV.
"Our browser-use script needs to move to Stagehand"	`browser-use-to-stagehand` maps the Python script onto Stagehand v3 primitives and picks a determinism level per step instead of a straight transliteration.

Verification¶

The raw path for the core skill returned HTTP 200 on branch main when this guide was written on 2026-10-31:

curl -sL -o /dev/null -w "%{http_code} browser\n" \
  https://raw.githubusercontent.com/browserbase/skills/main/skills/browser/SKILL.md

Expect 200. After the copy lands, confirm the count and that scripts/ came along, since browser-trace, browser-to-api and webmcp-gen lose most of their value without it:

ls -d ~/.hermes/profiles/corpusiq/skills/*/ | wc -l
head -6 ~/.hermes/profiles/corpusiq/skills/browser/SKILL.md
ls ~/.hermes/profiles/corpusiq/skills/browser-trace/scripts

The count should print 18 on a profile that had no other skills, the header should show name: browser with a license: MIT line and a compatibility line naming the browse CLI, and the script listing should show the Node helpers the trace skill runs.

Security¶

One sampled skill carries a Fail and the repository ships no license: add-webmcp fails the Gen Agent Trust Hub auditor, and there is no LICENSE file at the root of the repository, so the suite is Yellow and both facts matter before an install.

skills.sh verdicts for sampled skills (verified 2026-10-31) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
browser	Pass	Pass	Warn
search	Pass	Pass	Warn
event-prospecting	Pass	Pass	Warn
webmcp-gen	Pass	Warn	Pass
add-webmcp	Fail	Pass	Pass

The warn surface is what the skills do rather than what they contain. Every skill is a markdown bundle, but ten of them ship scripts/ folders whose helpers shell out to the browse CLI, drive Chrome over CDP, or write trace directories under the working tree, so the risk is the commands an agent is told to run. cookie-sync deserves its own note: it reads the operator's local Chrome profile and copies authenticated cookies into a Browserbase context, which is the most sensitive operation in the suite and the one to scope deliberately. safe-browser is the opposite case, generating an agent with a single allowlisted browser tool and no shell access at all.

Two mechanical details are worth checking yourself. The root package.json declares "postinstall": "npm run build", so installing from the repository root runs a TypeScript build rather than copying files; a directory copy of skills/ avoids that entirely. And the missing LICENSE is real at the repository level even though each SKILL.md carries license: MIT in its frontmatter, so the per-skill declaration is the only grant text in the tree. Treat API keys and cookies as environment state, never as prompt content, and keep BROWSERBASE_API_KEY scoped to the session that needs it.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-10-31: 40,127 combined installs across 20 indexed listings; 3,747 GitHub stars; no LICENSE file; last pushed 2026-10-27. Counts drift over time.
The index runs ahead of the tree. Twenty listings are indexed but the clone holds 18 skill directories: browserbase-cli and what-antibot resolve to nothing under skills/, so an install of "everything" installs 18 and two table rows have no route to point at.
Install counts are not a ranking of usefulness. browser leads at 5,909 while optimize-agent-prompt sits at 229 and add-webmcp at 198, even though add-webmcp is the one skill in the sample with a security Fail; volume says what people tried, not what holds up.
There is no LICENSE file. The repository root has no grant, and the MIT declared in each SKILL.md frontmatter is a per-file claim rather than a repository licence, which is a question for legal review before this ships inside a product.
The skills are not self-contained. Most depend on the global browse CLI, remote sessions depend on BROWSERBASE_API_KEY, and browser-to-api depends on a trace directory produced by browser-trace, so a partial install can leave a skill that cannot run.
Several skills overlap on purpose. browser, safe-browser and autobrowse all claim browser automation, and ui-test and agent-experience both test software, so routing depends on the phrasing of the request rather than on a clean boundary.

Skills Catalog | Skills Marketplace

Pro Tips¶

Install the four-skill core first: browser, fetch, search and browser-trace. That covers navigation, cheap static fetches, discovery and debuggability, and it is the combination that makes every other skill in the repository easier to evaluate.
Reach for fetch or search before browser on any read-only task. Both run without a browser session, so they cost a request instead of a session, and the skills say so in their own descriptions.
Keep traces. browser-trace writes per-page buckets into a run directory, and browser-to-api and autobrowse both consume those directories; a run without a trace is a run the agent cannot learn from.
Pass --local on the first browser command for localhost work, and --auto-connect when the agent should reuse an existing Chrome session, both of which the README documents and neither of which is the default.
Skip the root npm install. Copy skills/ as directories so the postinstall build never runs, and read the add-webmcp skill itself before adopting it, given its Gen Agent Trust Hub Fail.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ