browse

--- name: browse description: Use the browse CLI for Browserbase browser automation, Browserbase cloud APIs, Browserbase Functions, templates, web fetch/search, diagnostics, and Browse.sh skill discovery/installation. Use when the user asks to navigate pages, inspect browser state, run local or remote browser sessions, manage Browserbase resources, call Browserbase Functions, browse or scaffold Browserbase templates, fetch or search web content, diagnose browse setup, find or install a skill for a website task, discover site-specific Browse.sh skills, or install/refresh this browse skill. compatibility: "Requires the browse CLI (`npm install -g browse`). Remote Browserbase sessions and cloud API commands require `BROWSERBASE_API_KEY`. Local mode uses Chrome/Chromium on the machine." license: MIT allowed-tools: Bash metadata: openclaw: requires: bins: - browse install: - kind: node package: browse bins: [browse] homepage: https://github.com/browserbase/stagehand/tree/main/packages/cli --- # Browse CLI Use `browse` as the primary Browserbase command-line interface. It can: - drive a local or Browserbase-hosted browser session - inspect pages through accessibility snapshots, screenshots, DOM/text reads, and network capture - interact with pages by refs, selectors, XPath, keyboard, mouse, files, and viewport controls - manage Browserbase projects, sessions, contexts, extensions, fetch, and search APIs - develop, publish, and invoke Browserbase Functions - browse and scaffold Browserbase templates - diagnose local or remote browser setup issues - discover and install Browse.sh catalog skills - install or refresh this Browse CLI skill ## Setup Check Verify the CLI exists before relying on it: ```bash which browse || npm install -g browse browse --help ``` Install or refresh this skill with: ```bash browse skills install ``` Use `browse <topic> --help` for exact flags before running unfamiliar commands. ## Browser Target Selection Browser driver commands auto-start the browse daemon when needed. Choose the browser target per command with flags: ```bash browse open https://example.com --local browse open https://example.com --local --headed browse open https://example.com --remote browse open https://example.com --remote --verified --proxies browse open https://example.com --auto-connect browse open https://example.com --cdp 9222 browse open https://example.com --cdp ws://127.0.0.1:9222/devtools/browser/<id> ``` Use local mode for development, localhost, trusted sites, and fast iteration. Use `--auto-connect` only when the user explicitly wants to attach to an already-running debuggable Chrome session with existing cookies or login state; use `--local` when no debuggable Chrome is available. Use remote mode when Browserbase credentials are available and the site needs hosted browser infrastructure, Verified browser mode, CAPTCHA solving, proxies, or session persistence. `--local` requires Chrome or Chromium already installed on the machine. In containers, CI, and sandboxes with no browser installed, use `--remote` instead of `--local`. If `--local` fails with "No Chrome or Chromium found" and `BROWSERBASE_API_KEY` is set, switch to `--remote` — do not retry `--local`. For a Verified and/or proxied remote session, add `--verified` and/or `--proxies` to `--remote` — a single command that keeps the Browserbase session identity, so `browse status` and `browse doctor` report the session ID and live-view URL. `--verified` requires a Browserbase Scale plan. These flags only apply to `--remote` and are sticky for the session's lifetime, like `--headed`. Reach for `browse cloud sessions create` + `--cdp` only when you need session options `open` doesn't expose (region, keep-alive, contexts). Choose headed/headless and local/remote mode when starting a session. A running session keeps its mode: passing a conflicting flag such as `--headed` to an already-running headless session fails until you run `browse stop --session <name>` or target a different session. Use named sessions for any non-trivial work, especially when multiple agents or parallel tasks may run at once. Every browser command accepts `--session <name>` (or `-s <name>`); the `BROWSE_SESSION` env var sets the default, and commands without either share the `default` session. If `BROWSE_SESSION` is already set in the environment, every command already targets that session — do not pass `--session` or invent a new name. An explicit `--session <name>` always overrides `BROWSE_SESSION` for that command, so only pass it to deliberately target a different session. ```bash browse open https://example.com --session research --local browse snapshot --session research ``` Remote browser and cloud API commands require: ```bash export BROWSERBASE_API_KEY=... ``` The CLI forwards `BROWSERBASE_API_KEY` from the caller on every browser command, even when the daemon is already running. If a remote command first fails because the key was missing, set or export the key and retry the command; do not restart the daemon just to refresh the environment. ## Browser Automation Workflow Start by opening the page, then inspect state, act, and verify. ```bash browse open https://example.com --session research --local browse snapshot --session research browse click @0-5 --session research browse type "hello" --session research browse snapshot --session research browse stop --session research ``` Prefer `browse snapshot` over screenshots for most browser work. It is structured, fast, and returns refs like `@0-5` for reliable element interaction. Use screenshots when visual layout, images, or pixel-level state matter. Refs are refreshed on every snapshot. After clicks, form submits, navigation, or UI re-renders, take a new snapshot before using another ref. ## Parallel Browser Work Use a different `--session` value for each independent browser task. Sessions isolate tabs, cookies, refs, and daemon state; parallel tasks that omit `--session` share the `default` session and overwrite each other's active page. ```bash browse open https://example.com/search-a --session search-a --local browse open https://example.com/search-b --session search-b --local browse snapshot --session search-a browse snapshot --session search-b ``` When a task is complete, stop only that task's session: ```bash browse stop --session search-a ``` ## Core Browser Commands Navigation: ```bash browse open <url> browse reload browse back browse forward browse wait load browse wait selector "#result" ``` Page state: ```bash browse snapshot # formatted tree only browse snapshot --full # also include ref maps (xpathMap, urlMap) browse get url browse get title browse get text body browse get html "#main" browse get value "#email" browse get markdown body # page/element content as markdown browse eval "document.title" # run JavaScript in the active page browse screenshot # saves screenshot-<timestamp>.png, prints { "saved": "<path>" } browse screenshot --path page.png # choose the output path browse screenshot --base64 # legacy: print base64 JSON to stdout (avoid in agent loops) ``` Interaction: ```bash browse click @0-5 browse fill @0-8 "search query" browse type "text for the focused element" browse press Enter browse select "select[name=country]" "United States" browse upload @0-12 ./file.pdf browse highlight @0-5 browse is visible "#modal" ``` Mouse and viewport: ```bash browse mouse click 240 320 browse mouse hover 240 320 browse mouse drag 80 80 310 100 browse mouse scroll 500 300 0 600 browse viewport 1280 720 browse cursor # show a visible cursor overlay ``` Tabs, network, and CDP: ```bash browse tab list browse tab new https://example.com browse tab switch <target-id> browse tab close <target-id> # refuses to close the last tab browse network on browse network off browse network path browse network clear browse cdp 9222 --pretty ``` Session management: ```bash browse doctor browse doctor --json browse status browse stop browse stop --force ``` Use `browse doctor` before debugging a broken browser session. Use `browse doctor --json` when another agent or CI needs structured diagnostics. `browse stop` is idempotent. If no daemon is running, `{ "stopped": false }` with exit code `0` is successful cleanup, not an error. If a page command reports that no active page is available, inspect and recover the named session: ```bash browse status --session research browse tab list --session research browse tab new https://example.com --session research browse open https://example.com --session research ``` ## Cloud APIs Use `browse cloud` for Browserbase platform APIs: ```bash browse cloud projects list browse cloud projects get <project-id> browse cloud projects usage <project-id> browse cloud sessions create browse cloud sessions create --proxies --verified browse cloud sessions list browse cloud sessions get <session-id> browse cloud sessions update <session-id> browse cloud sessions debug <session-id> browse cloud sessions logs <session-id> browse cloud sessions downloads get <session-id> browse cloud sessions uploads create <session-id> ./file.pdf browse cloud contexts create --name github browse cloud contexts add github <context-id> browse cloud contexts list browse cloud contexts get <context-id|name> browse cloud contexts update <context-id|name> browse cloud contexts delete <context-id|name> browse cloud extensions upload ./extension.zip browse cloud extensions get <extension-id> browse cloud extensions delete <extension-id> browse cloud fetch https://example.com browse cloud search "browser automation" ``` For remote sessions with context persistence: ```bash browse cloud sessions create --context-id <context-id> --persist ``` Contexts persist cookies and local storage (logins) across sessions. Name a context once with `--name` to save a local alias, then reuse the name anywhere a context ID is accepted instead of memorizing the ID: ```bash browse cloud contexts create --name github # saves github -> ctx_... browse cloud contexts add github <context-id> # name a context you already have browse cloud sessions create --context-id github --persist browse cloud contexts list # show saved names ``` Use `--verified` when the task needs Browserbase Verified browser mode. To drive a Verified/proxied session directly, prefer `browse open <url> --remote --verified --proxies` over create-then-attach — it keeps the session identity so `browse status`/`browse doctor` can report it. Use `browse cloud sessions create` for session options the driver flags don't cover (region, keep-alive, contexts, full `--stdin` body). Use `browse cloud fetch` when the user needs a simple HTTP fetch without browser interaction. It returns markdown-formatted page content by default; pass `--format raw` for the original response body or `--format json --schema <schema>` for structured extraction. Use `browse cloud search` when the user asks for web search results. ## Browserbase Functions Use `browse functions` to create, develop, publish, and invoke Browserbase Functions: ```bash browse functions init my-function browse functions dev index.ts browse functions publish index.ts browse functions publish index.ts --dry-run browse functions invoke <function-id> --params '{"url":"https://example.com"}' browse functions invoke --check-status <invocation-id> ``` Functions commands use `BROWSERBASE_API_KEY`. Generated projects import `defineFn` from `@browserbasehq/sdk-functions`. ## Templates Use `browse templates` to discover and scaffold Browserbase starter templates: ```bash browse templates list browse templates list --tag Python --source Browserbase browse templates find google-trends-keywords browse templates find amazon --json browse templates clone google-trends-keywords browse templates clone amazon-product-scraping --language python ./my-scraper browse templates clone dynamic-form-filling ./form-bot --language typescript ``` Use `browse templates find` before cloning when the exact slug is uncertain. Use `--language typescript` or `--language python` to choose the generated project runtime when a template supports both. ## Skills Install or refresh this bundled CLI skill: ```bash browse skills install ``` ### Discovering Browse.sh Skills Browse.sh (https://browse.sh) is a catalog of site-specific browser automation skills. Each skill is scoped to one task on one website and identified by a `<domain>/<task>` slug. An installed skill encodes a proven strategy for that site — API endpoints, selectors, anti-bot handling — so it completes the task faster and more reliably than exploring the site from scratch. Search the catalog proactively when: - the user asks to complete a task on a specific website — search the domain before automating it by hand - the user asks for a task in a common category like flights, food delivery, reviews, recipes, tickets, jobs, or shopping — search the keyword - the user asks "is there a skill for X" or wants new capabilities ```bash browse skills list # browse the catalog browse skills find <query> # search by slug, domain, title, description, category, alias, or tag browse skills add <domain>/<task> # install an exact slug ``` ### Search Strategy Search the domain first, then the task, then broaden: ```bash browse skills find yelp.com # 1. exact domain browse skills find yelp # 2. site name browse skills find reviews # 3. task keyword browse skills find "restaurant reviews" # 4. multi-word query browse skills find food --limit 5 # 5. category keyword, capped ``` - Querying an exact slug (`browse skills find yelp.com/extract-reviews-2ikb22`) prints a detail view with the full description and install command. - If a search returns nothing, try synonyms (`flights` vs `travel`, `food` vs `restaurants`) before concluding no skill exists. - Each result shows a recommended method — `api`, `fetch`, `browser`, or `hybrid` — indicating how the skill drives the site. Install counts signal which skills are proven. - Many slugs end in a generated suffix (`-2ikb22`), so never guess or construct them. Install only with an exact slug copied from `list` or `find` output: `browse skills add yelp.com/extract-reviews-2ikb22`. The installed skill becomes available to the agent as a regular skill; a new agent session may be needed to pick it up. ### Output Formats Output is a table in a terminal and JSON when piped, so agents get structured JSON by default. Force a format with `--format table|json` or `--json`. ```bash browse skills find food --format table --limit 10 browse skills find food --json | jq -r '.skills[] | "\(.slug)\t\(.title)"' ``` JSON output includes every match with full descriptions and ignores `--limit`; `browse skills list --json` returns the entire catalog, which is large. Prefer `browse skills find <query>` to narrow first, or `--format table --limit <n>` for a compact view. ## Best Practices 1. Run the real command and inspect its output instead of guessing. 2. Use `browse snapshot` before interacting so you have current refs. 3. Re-run `browse snapshot` after navigation or DOM-changing actions because refs can change. 4. Prefer refs from snapshots for clicks and uploads; use selectors or XPath when refs are unavailable. 5. Use `--local` for localhost and repeatable development; use `--remote` for protected sites or Browserbase-specific behavior. 6. Use a distinct `--session <name>` for each parallel or long-running task; commands without the flag share the `default` session. 7. Use `--auto-connect` only when attaching to an existing debuggable local Chrome session is intended. 8. Use `browse doctor` when session startup, browser discovery, CDP attach, or Browserbase auth looks wrong. 9. Never retry a failing command unchanged. If the same command fails twice with the same error, stop — run `browse doctor --json`, then change approach (fix the key, switch `--local`/`--remote`, or `browse stop --force` and start fresh). Repeating an identical failing command will keep failing. 10. Use `browse stop` (or `browse stop --session <name>`) when finished to clean up daemon state. 11. For unfamiliar command details, run `browse <topic> --help` and follow the exact dash-case flags. ## Troubleshooting - "No active page": run `browse status --session <name>`, then `browse open <url> --session <name>` or `browse tab new <url> --session <name>`; use `browse stop --force` if the daemon is stale. - "Driver daemon ... is not running": run the exact `browse open` command printed by the error. Browser commands start the daemon automatically. - Chrome not found: use `--remote` with Browserbase credentials, install Chrome, or attach with `--cdp`. - Action fails: run `browse snapshot` and use a visible ref from the current page state. - 401 Unauthorized on `open`, `get`, or other driver commands: a set `BROWSERBASE_API_KEY` makes `browse` default to remote mode. Fix the key at https://browserbase.com/settings, unset it, or pass `--local` to run a managed local browser (no key needed). - Same command fails twice with the same error: stop retrying — never retry a failing command unchanged. Init failures are cached for several seconds, so instant retries return identical errors. Run `browse doctor --json`, then change approach: fix the key, switch `--local`/`--remote`, or `browse stop --force` and start fresh. - Remote command fails: verify `BROWSERBASE_API_KEY` and inspect `browse cloud projects list`. - Session setup is unclear: run `browse doctor` or `browse doctor --json`. - Protected site blocks local mode: retry with `--remote`. - `browse skills find` returns nothing: broaden the query — bare domain, then site name, then a task keyword or synonym. - `browse skills add` fails on `npx`: install Node.js from https://nodejs.org, then rerun. - Newly added skill is not available: it installs for future sessions; list installed skills or start a new agent session.