메뉴
Drive a browser via CDP — navigate, click, type, screenshot. Profiles persist login state across sessions, so agents log in once and stay authenticated. Triggers on: 'automate a website', 'browser automation', 'click', 'screenshot', 'scrape', 'log in', 'fill form'.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Gotchas for the agents-cli headless release. The macOS home base signs, notarizes, and npm-publishes over SSH with no GUI and no Touch ID — three things must be primed or the release stalls. Triggers on: release stuck at sign/publish, errSecInternalComponent, "no npmjs.com bundle", "Missing embedded.provisionprofile", making the release autonomous.
Search, browse, and read agent conversation transcripts across Claude, Codex, Gemini, and OpenCode. Use this skill to find previous sessions, recover context, or inspect what agents have done.
Organize AI coding agents into teams that collaborate on a shared task. Create teams, add teammates, start them, monitor progress, and collect results. Use this skill when you need parallel agent execution. For single-agent dispatch, use `agents run` instead.
Use FIRST for any question about THIS codebase's structure, symbols, or relationships — where a function/class/file is defined, what calls or imports it, what breaks if you change it, how two parts connect, or to load focused source context before editing or reviewing. Routes to the cgraph MCP tools (graph_query / graph_explain / graph_impact / graph_path / graph_context), which serve ranked file:line results, node neighborhoods, transitive blast radius, and token-budgeted source bundles from a resident per-project graph daemon in ~10ms — far cheaper than grepping and reading files. Prefer over blind grep/read for code navigation, dependency tracing, and impact analysis.
Register and connect to your machines over Tailscale SSH with agents-cli. Use this skill to sync devices from the tailnet, list them, open a shell on another machine, or see agent sessions running across your whole fleet.
Schedule agents to run on a cron schedule or one-shot at a specific time. The scheduler auto-starts on first add. Triggers on: 'schedule an agent', 'recurring job', 'cron', 'daily', 'every weekday', 'run at 2pm'.
| name | browser |
| description | Drive a browser via CDP — navigate, click, type, screenshot. Profiles persist login state across sessions, so agents log in once and stay authenticated. Triggers on: 'automate a website', 'browser automation', 'click', 'screenshot', 'scrape', 'log in', 'fill form'. |
| argument-hint | [profiles|start|refs|click|type|screenshot|done] |
| allowed-tools | Bash(agents browser*), Bash(browser*) |
| user-invocable | true |
Control a real browser via CDP. Profiles persist login state, so agents log in once and stay authenticated. Screenshots auto-resize to save tokens.
browser is shorthand for agents browser — use the short form.
Every action command targets a task. Bind the task once per shell, then drop the flag:
export AGENTS_BROWSER_TASK=$(browser start --profile work)
browser refs
browser click 42
browser screenshot
start writes the resolved task name (e.g. swift-crab-falcon-a3f92b1c) to stdout; human commentary goes to stderr — $(...) capture stays clean.
For per-call overrides, pass --task <name>. Env vars are per-process, so parallel agents in different shells never collide.
Use your real browser. Same fingerprint, same IP, nothing to detect:
browser profiles create work --browser chrome
export AGENTS_BROWSER_TASK=$(browser start --profile work --url https://linkedin.com)
browser refs
browser click 5
Log in once to a profile — the session persists. Every future task is already authenticated:
browser profiles create social --browser chrome
export AGENTS_BROWSER_TASK=$(browser start --profile social)
# Log in manually or via automation
browser done
# Next time — already logged in
export AGENTS_BROWSER_TASK=$(browser start --profile social --url https://twitter.com)
Profiles are identities. Create one per account group:
browser profiles create social --browser chrome # Twitter, LinkedIn
browser profiles create email --browser chrome # Gmail, work email
browser profiles create finance --browser chrome # Banking, trading
An agent using social can't see your finance cookies.
Each agent gets its own task. Tasks share the window but own separate tabs:
# Agent 1 — research
export AGENTS_BROWSER_TASK=$(browser start --profile work --url https://arxiv.org)
browser tabs
# Agent 2 (different shell) — monitoring
export AGENTS_BROWSER_TASK=$(browser start --profile work --url https://dashboards.example.com)
browser tabs # only sees its own tabs
# Completing one doesn't affect the other
browser done # closes only this agent's tabs
Cookie persistence comes first. A profile keeps its login cookies across sessions, so the goal is: log in once (by hand or agent-driven), and the profile stays authenticated — no credential handling on the hot path. Check state before assuming a re-login is needed:
browser profiles logins # per profile: SERVICE | ACCOUNT | CREDS
ACCOUNT is the signed-in identity (from saved logins, read plaintext — no decryption);
CREDS shows whether the profile's secrets bundle holds login creds for that service.
When a session lapses, drive the login with the credential resolved inside the
browser layer so the plaintext never crosses stdout or your transcript. Store creds in
the profile's bundle under the <SERVICE>_USERNAME / <SERVICE>_PASSWORD convention:
browser profiles create acme --browser comet --secrets acme-login
agents secrets add acme-login GITHUB_USERNAME
agents secrets add acme-login GITHUB_PASSWORD
# Log in — the agent drives the form and handles 2FA/selectors adaptively:
export AGENTS_BROWSER_TASK=$(browser start --profile acme --url https://github.com/login)
browser refs # find the field refs
browser type <user-ref> --secret acme-login/GITHUB_USERNAME # value never printed
browser type <pass-ref> --secret acme-login/GITHUB_PASSWORD
browser click <submit-ref>
browser screenshot # inspect: 2FA? captcha? done?
If a 2FA/OTP or "unusual activity" checkpoint appears, the agent screenshots it and asks
the human (or reads the OTP from a connected mailbox) — the CLI never auto-solves it.
For scripts that genuinely need the raw value, agents secrets get <bundle> <KEY> prints
one value (audited) — but prefer type --secret so nothing lands in the transcript.
Note: --secrets also injects the bundle as env vars into the browser process at
launch (useful for extensions/CDP tooling that read env) — that is separate from web
login, which needs the form-fill flow above.
Connect to BrowserBase, Steel, or any CDP service:
browser profiles create cloud --browser chrome \
--endpoint "wss://connect.browserbase.com?apiKey=..."
Three flags work together for Electron desktop apps:
browser profiles create canva --browser custom \
--binary "/Applications/Canva.app/Contents/MacOS/Canva" \
--electron \
--target-filter "url:https://www.canva.com/"
--browser custom plus --binary tells the launcher how to start the app with --remote-debugging-port.--electron switches the runtime into single-window mode: tabs are never created with Target.createTarget (most production Electron apps reject it); navigation reuses the existing window.--target-filter picks the visible WebContents. Electron apps frequently expose several type: page CDP targets — background services, OAuth windows, and file:// shells — and the first one CDP returns is almost never the UI. Use url:<substring> or title:<substring>.Without --target-filter, a skip-invisible heuristic excludes about:blank, file://, and URLs matching _desktop-background-service / _internal / _background. That covers most apps; use the filter for the ones it doesn't.
export AGENTS_BROWSER_TASK=$(browser start --profile work --url https://example.com)
browser refs # Get clickable elements
browser click 3 # Click element ref 3
browser type 5 --text "search query" # Type into element ref 5
browser press Enter # Press Enter key
browser screenshot # Take screenshot (auto-saved)
browser done # Close task's tabs
browser tab add --url https://github.com # Open new tab (becomes current)
browser tab focus github # Switch by URL substring
browser tabs # List all tabs
browser tab close # Close all tabs
browser status # Show all profiles and tasks
browser tasks # List just tasks
All action commands read the task from $AGENTS_BROWSER_TASK unless --task <name> is passed.
| Task | Command |
|---|---|
| Create profile | browser profiles create <name> --browser chrome |
| Start task | browser start --profile <name> [--url <url>] |
| Navigate | browser navigate --url <url> |
| Get elements | browser refs |
| Click | browser click <ref> |
| Type | browser type <ref> --text "text" |
| Press key | browser press Enter |
| Hover | browser hover <ref> |
| Scroll | browser scroll --dx 0 --dy 1000 (negatives scroll up/left) |
| Screenshot | browser screenshot |
| Evaluate JS | browser evaluate --expression "document.title" (or --file ./script.js) |
| New tab | browser tab add --url <url> |
| Switch tab | browser tab focus <hint> |
| Complete task | browser done |
| Status | browser status |