Menu
Chrome browser control: open pages, take ref snapshots, click, type, screenshot. Requires cli-jaw server running.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
MUST USE for cli-jaw PABCD orchestration workflows — orchestrate, phase, attest/attestation, interview mode, goal mode, checkpoints, and multi-phase development. Triggers: orchestrate, phase, attest, attestation, interview, goal mode, checkpoint, PABCD, 요구사항 정리, 인터뷰, 스펙 정리. Operate state transitions only when the user explicitly requests orchestration or an active PABCD phase is injected — do not transition state merely because a document mentions phases, goals, or checkpoints.
MUST USE for every coding task — classifies work depth (C0-C5), task_tags overlays, modular limits, pre-write search, verification-before-completion, and safety rules. Triggers: develop, implement, refactor, feature, bug fix, test, review, code quality, scaffolding.
Use only on the Codex CLI for native image generation or image editing without an API key. Save final PNG files under ~/.cli-jaw/uploads, report web-ready absolute-path markdown, and send to Telegram or Discord only when explicitly requested.
MUST USE for any real runtime debugging in any language — crashes, silent failures, wrong output, build/test failures, flaky tests, performance regressions, and integration bugs. A 5-phase root-cause method: architecture check → investigate → analyze → hypothesize → implement. Triggers: debug this, why is X failing, flaky test, fix the crash, root cause, error, stack trace, regression, 왜 안 돼, 디버깅, 원인 분석.
MUST USE for code review and review-readiness — review process, quality thresholds, antipattern detection, review verdicts, and giving/receiving feedback. Triggers: review this, code review, PR review, check my diff, before merge, antipattern, review-readiness, 리뷰, 코드 리뷰, 머지 전에 확인.
MUST USE for any frontend, web UI, or visual implementation work — building, styling, or redesigning pages/components, responsive layouts, motion, component architecture, and production-surface polish. Triggers: frontend, UI, component, CSS, responsive, animation, React, Vue, Svelte, Tailwind, layout, styling, redesign, mockup, anti-slop, 프론트엔드, UI 작업, 반응형, 디자인 수정.
| name | browser |
| description | Chrome browser control: open pages, take ref snapshots, click, type, screenshot. Requires cli-jaw server running. |
| metadata | {"openclaw":{"emoji":"🌐","requires":{"bins":["cli-jaw"],"system":["Google Chrome"]},"install":[{"id":"brew-cliclick","kind":"brew","formula":"cliclick","bins":["cliclick"],"label":"Install cliclick (optional, for coordinate-based clicks)"}]}} |
Control Chrome through cli-jaw browser commands.
Use ref-based snapshots to identify page elements, then click/type by ref ID.
Role separation: the search skill discovers and routes (queries → URL
candidates); this browser skill verifies evidence when fetch/snippets are not
enough (original page, DOM, PDF, tables). Search finds, browser proves.
This skill follows the newer 30_browser workflow shape, adapted for the
server-backed cli-jaw browser runtime. Commands that are not implemented in
the current cli-jaw runtime are separated under Planned Runtime Delta and
must not be used as current commands.
cli-jaw serve must be running.playwright-core must be installed in the cli-jaw project.cli-jaw browser start --agent # Automation session (headless, no visible test window)
cli-jaw browser start # Interactive browser (manual only)
cli-jaw browser start --headless # Manual headless mode (server/CI/WSL)
cli-jaw browser navigate "https://example.com" # Go to URL
cli-jaw browser snapshot --interactive # Interactive elements with ref IDs
cli-jaw browser click e3 # Click ref e3
cli-jaw browser type e5 "hello" --submit # Type + Enter
cli-jaw browser screenshot # Save screenshot path
Always follow this pattern:
snapshot --interactive -> act by ref or key -> snapshot -> verify
Use a fresh snapshot after navigation, reload, tab changes, or any action that substantially changes the page. Ref IDs belong to the latest usable snapshot and can go stale.
The default browser lane is standalone Chrome via cli-jaw browser (above).
Separately, the Electron Manager has an EMBEDDED browser (right-sidebar
Browser tab). You cannot open or enumerate it yourself — it becomes available
ONLY when the user clicks "Share with Agent" on a page. A shared page appears
in your runtime-context as an [Embedded Browser] entry with a target id and
exact curl commands (correct manager port included).
POST .../api/manager/embedded-browser/<targetId>/screenshot # PNG temp-file path
POST .../api/manager/embedded-browser/<targetId>/snapshot # bounded accessibility tree (roles/names/bounds)
POST .../api/manager/embedded-browser/<targetId>/act # click/type/scroll/key
Rules:
act needs a SECOND user opt-in ("Allow agent actions" toolbar toggle) —
the runtime-context entry says whether actions are enabled. If not enabled,
stay read-only (screenshot/snapshot) and ask the user to act or to enable
actions.act payloads: {"act":{"kind":"click","x":..,"y":..}},
{"kind":"type","text":".."}, {"kind":"scroll","x":..,"y":..,"deltaY":..},
{"kind":"key","key":"Enter"}. Use snapshot bounds for coordinates;
re-snapshot after navigation or layout changes.cli-jaw browser Chrome lane.These commands are implemented in the current cli-jaw browser runtime.
| Surface | Label | Notes |
|---|---|---|
local cli-jaw browser primitives | ready | server-backed local Chrome/CDP only |
doctor and cleanup-runtimes | ready | dry-run by default; close requires --force |
| dashboard visible/headless start split | ready | visible manual and headless agent modes are separate |
| web-ai provider workflows | beta | use the web-ai skill and provider-specific gates |
| external hosted/cloud CDP | deferred | do not claim remote browser hosting support |
cli-jaw browser start [--port <auto>] [--headless] [--agent]
cli-jaw browser stop
cli-jaw browser status
cli-jaw browser doctor [--json]
cli-jaw browser cleanup-runtimes [--json] [--close --force]
cli-jaw browser reset [--force]
--agent enables an automated headless session.browser start is for user-requested interactive browsing.doctor reports CDP/runtime ownership mismatch and orphan cleanup scope.cleanup-runtimes is dry-run by default; it only closes durable jaw-owned
orphan runtime records when both --close and --force are supplied.reset clears the browser profile and screenshots; use only when the user
explicitly wants a reset or you have confirmed it.cli-jaw browser snapshot
cli-jaw browser snapshot --interactive
cli-jaw browser snapshot --interactive --max-nodes 30 --json
cli-jaw browser screenshot
cli-jaw browser screenshot --full-page
cli-jaw browser screenshot --ref e5
cli-jaw browser screenshot --json
cli-jaw browser screenshot --clip 0 0 320 180 --json
cli-jaw browser text
cli-jaw browser text --format html
cli-jaw browser get-dom --selector ".card" --max-chars 2000 --json
cli-jaw browser console --json --limit 20
cli-jaw browser network --json --limit 20
e1 link "Gmail"
e2 link "Images"
e3 textbox "Search" <- To type here: type e3 "query"
e4 button "Google Search" <- To click: click e4
e5 button "I'm Feeling Lucky"
cli-jaw browser click e3
cli-jaw browser click e3 --double
cli-jaw browser click e3 --right
cli-jaw browser type e3 "hello"
cli-jaw browser type e3 "hello" --submit
cli-jaw browser press Enter
cli-jaw browser press Escape
cli-jaw browser press Tab
cli-jaw browser hover e5
cli-jaw browser mouse-click 400 300
cli-jaw browser mouse-click 400 300 --double
cli-jaw browser select e7 "option1"
cli-jaw browser drag e3 e5
cli-jaw browser move-mouse 400 300
cli-jaw browser mouse-down
cli-jaw browser mouse-up --right
cli-jaw browser navigate "https://example.com"
cli-jaw browser open "https://example.com"
cli-jaw browser tabs
cli-jaw browser tabs --json
cli-jaw browser active-tab --json
cli-jaw browser tab-switch 2
cli-jaw browser reload
cli-jaw browser resize 1440 900
cli-jaw browser scroll --x 0 --y 1000
cli-jaw browser wait-for-selector ".toast-success" --timeout 30000
cli-jaw browser wait-for-text "Dashboard" --timeout 30000
cli-jaw browser evaluate "document.title"
evaluate is a top-level browser diagnostic command. Do not expose arbitrary
user-provided JavaScript through higher-level vendor workflows such as web-ai.
For ChatGPT web-ai workflows, use the web-ai skill. The browser skill owns
primitive page control; web-ai owns structured question rendering, active-tab
safety, and response baseline handling.
Use browser commands as downstream evidence checks after the search skill has produced URL candidates. Search snippets and AI summaries are not final evidence.
Recommended ladder:
cli-jaw browser fetch "<url>" --json
cli-jaw browser open "<url>"
cli-jaw browser text
cli-jaw browser snapshot --interactive
cli-jaw browser get-dom --selector "<selector>" --max-chars 4000 --json
cli-jaw browser network --json --limit 40
Escalate through the ladder when the candidate URL is important and the current evidence is weak:
On browser escalation, fetch also runs an in-page Defuddle pass that
extracts the main content as markdown (tables, links, and footnotes
preserved). When the JSON evidence includes browser-defuddle, the returned
content is that markdown extraction — prefer it over raw page text for
tables/lists and X/article pages. If it fails (strict CSP), fetch degrades to
plain text and records a defuddle:* warning.
For Korean public/current searches, preserve source-sensitive status in the
answer: sufficient only after original evidence is visible, browse-needed
when browser escalation is still required, and insufficient when the source
cannot be reached.
cli-jaw browser fetch <url> is the known-URL reader lane. It can read a direct
candidate URL or a search-result URL, but it is not generic search and must not
receive a raw natural-language query.
Use this ladder for public-source reading (each step triggers only when the previous returned blocked, empty, or low-quality content):
r.jina.ai prefix reader for clean markdown
extraction with JS rendering. Enabled by default (--allow-third-party-reader);
disable with --no-allow-third-party-reader. 429 triggers a 60s cooldown.pip install camoufox[geoip]). Falls back silently when absent.structured-extractor.ts. Available alongside
defuddle main-content extraction.get-dom, selector-bound
reads, snapshots, and screenshots when visual or interactive evidence matters.Optional media reader: yt-dlp integration for YouTube (and 1,800+ sites)
metadata and transcript extraction. Emits ytdlp source candidate alongside
oembed when yt-dlp binary is detected. Falls back when absent.
Adaptive-fetch keeps live public-site smoke targets as a default-off manifest.
Use it for drift checks only when explicitly running verification; ordinary unit
tests must not hit live sites. Optional media helpers such as captions or
yt-dlp-style extraction remain opt-in and must not auto-install dependencies.
Stop rather than bypass when the page requires login, payment, private
membership, user credentials, or CAPTCHA solving. Report browse-needed when a
candidate likely needs browser/human verification and insufficient when no
credible public route remains.
When the user explicitly wants to drive a single Chrome instance (for
example: keep one logged-in profile open, avoid running both cli-jaw serve
and a second CDP session), the same browser commands are available through
the standalone agbrowse CLI (npm install -g agbrowse). The flag surface is
identical; only the binary prefix changes.
cli-jaw browser form | agbrowse form |
|---|---|
cli-jaw browser start --agent | agbrowse start |
cli-jaw browser status | agbrowse status |
cli-jaw browser navigate "<url>" | agbrowse navigate "<url>" |
cli-jaw browser snapshot --interactive | agbrowse snapshot --interactive |
cli-jaw browser click e3 | agbrowse click e3 |
cli-jaw browser type e5 "hello" --submit | agbrowse type e5 "hello" --submit |
cli-jaw browser screenshot | agbrowse screenshot |
cli-jaw browser tabs | agbrowse tabs |
cli-jaw browser stop | agbrowse stop |
Only switch when the user explicitly asks for the standalone path. For search
planning, agbrowse research plan is optional and does not replace native
cli-jaw search/browser verification. Do not run cli-jaw browser and agbrowse
against the same --port simultaneously —
the second start will reuse the first CDP and the persisted state files can
collide. For the web-ai layer, see the corresponding Standalone agbrowse Alternative section in the web-ai skill.
cli-jaw browser start --agent
cli-jaw browser navigate "https://www.google.com"
cli-jaw browser snapshot --interactive
cli-jaw browser type e3 "search query" --submit
cli-jaw browser snapshot --interactive
cli-jaw browser click e7
cli-jaw browser snapshot --interactive
cli-jaw browser type e1 "John Doe"
cli-jaw browser type e2 "john@example.com"
cli-jaw browser click e3
cli-jaw browser snapshot
cli-jaw browser navigate "https://news.ycombinator.com"
cli-jaw browser text
cli-jaw browser text --format html
cli-jaw browser snapshot --interactive
The copied 30_browser reference documents a richer command surface. These are
planned cli-jaw browser parity targets, not current commands unless the runtime
has been upgraded in a later PRD.
cli-jaw browser console --clear --reload --duration 3000
cli-jaw browser network --reload --duration 1000
cli-jaw browser wait 2000
cli-jaw browser resize 0 0 --fullscreen
cli-jaw browser scroll down
cli-jaw browser scroll up --amount 1000
wait-for <ref> is deprecated in the reference design because refs are
snapshot-scoped. Prefer selector/text waits.
If something goes wrong, stop and inspect state before the next action.
snapshot fails -> take screenshot for visual inspection.snapshot --interactive; refs can go stale.wait-for-selector or wait-for-text.status; only
stop/start when that is the selected recovery path.reset unless the user already
requested destructive reset.vision-click skill only after confirming no
usable ref exists.| Variable | Description |
|---|---|
CHROME_HEADLESS=1 | Enable headless mode for manual starts. |
CHROME_NO_SANDBOX=1 | Disable Chrome sandbox for Docker/CI only. |
The default CDP port is derived from the cli-jaw server port. Use
cli-jaw browser start --port <port> only when you need an explicit override.
cli-jaw browser start --headless
cli-jaw browser start --agent
CHROME_HEADLESS=1 cli-jaw browser start
Use --agent for automation. It avoids popping a visible browser window.
cli-jaw browser doctor --json
cli-jaw browser cleanup-runtimes
cli-jaw browser cleanup-runtimes --close --force
cleanup-runtimes is intentionally conservative. It only acts on a durable
browser-runtime-owner.json record written by jaw-owned Chrome launches, and
the process command line must still match the recorded pid, CDP port, and
profile. Chrome helper processes containing --type= are rejected. Never use
general ps output as proof that a Chrome process is safe to close.
| Symptom | Cause | Fix |
|---|---|---|
| CDP connection refused | Chrome not started or wrong port | cli-jaw browser status, then start with the expected port |
running:false with owner:jaw-owned | stale runtime metadata or dead CDP | cli-jaw browser doctor, then retry browser start |
| old headless jaw Chrome remains | durable jaw-owned orphan candidate | cli-jaw browser cleanup-runtimes dry-run before --close --force |
| Windows only opens test browser | Chrome singleton absorbed launch | Close all Chrome windows, then use start --agent |
| Headless CDP not opening | headless not requested in GUI-less env | Add --headless or use --agent |
| Port conflict | another process owns the CDP port | choose a different --port |
| Snapshot too large | page has many nodes | planned: --max-nodes; current: use --interactive |
snapshot --interactive after navigation or major page changes.--interactive for token budget.~/.cli-jaw/screenshots/.start --agent should be the default for agent automation.vision-click skill only as an explicit fallback.