
browser-use
Control the user's existing signed-in Chrome: callable Codex Chrome plugin first, then OpenClaw extension-backed mcporter, with direct DevTools attachment only as a last fallback.
- Installs
- 191
- Last 24h
- —
- Trending rank
- —
- Security
- Not audited
Install
Works with every agent the skills CLI supports.
npx skills add https://github.com/steipete/agent-scripts --skill browser-use- 01
Install
Run the command in your project. The CLI asks which agents to install for.
- 02
Describe the task
Ask your agent for something this skill covers. It loads browser-use automatically when relevant.
- 03
Review the result
The agent follows the SKILL.md steps, scripts and references bundled with the skill.
SKILL.md
The instructions your agent reads when the skill is triggered.
Browser Use
Control the user's existing real Chrome profile, especially for login-dependent work and live UI verification.
Route
- Use the Codex
ChromeorChrome [Internal]plugin when it is callable in the active session. Installed on disk is not enough. - Otherwise prefer the OpenClaw extension-backed mcporter route.
- Use legacy direct DevTools attachment only as the explicit last fallback.
For mcporter, the MCP call remains the agent-facing control interface. The
OpenClaw extension is the transport underneath it, not a separate tool. Set
chromeDevtoolsRelay: "require" on the canonical global Chrome server definition
to enforce relay-only routing. Keep the caller policy explicit as well:
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.<tool>
Seeing an agent request the chrome-devtools MCP tool is therefore expected.
A Chrome Allow remote debugging? prompt or a relay-policy error indicates
that the extension transport was not used. The shared daemon retains canonical
policy; a shell environment prefix alone cannot change that policy.
Never use isolated Chrome, the Codex in-app browser, Playwright, Puppeteer,
AppleScript, osascript, generic GUI scripting, or macOS open as a browser-control
substitute unless the user explicitly asks for an isolated or new browser.
Peekaboo is allowed only for Chrome or extension setup and visible prompts.
Login-heavy sites often depend on the real profile's cookies, SSO, device trust,
and extensions.
For a rendered-browser bug, prove behavior through this real profile. Treat
curl, source inspection, API checks, and isolated test browsers as supporting
evidence, not substitutes for live UI proof.
Extension Relay Model
OpenClaw creates a random per-host relay key in its mode-0600 credentials
directory. The extension and same-host clients use nonce-bound mutual HMAC
proofs. The reusable key is never sent to an unverified loopback listener,
placed in a URL, or passed to the child MCP process. Keep credentials out of
configuration, command output, chat, logs, and screenshots.
On a same-host relay, mcporter authenticates /json/version and upgrades the
same retained socket to /cdp, then gives chrome-devtools-mcp a protected
one-use local handoff. Agents still call the standard MCP tools; successful
relay routing is what removes direct DevTools attachment and Chrome's approval
prompt.
New pairings default to All tabs: every ordinary eligible tab is exposed except tabs explicitly paused in the popup. Existing pairings keep their stored mode. In Selected tabs mode, membership in the Chrome tab group titled OpenClaw is the sharing boundary. Restricted/internal pages, incognito, other profiles, and tabs without an eligible URL remain excluded in both modes.
Topology Boundary
Direct remote Gateway pairing over wss:// lets OpenClaw's Gateway-side
browser tool control local Chrome. It does not create a local relay for a local
mcporter process. Do not routinely copy remote secrets or build ad-hoc SSH
tunnels around this boundary.
If local mcporter cannot authenticate to a local relay, the extension-backed mcporter route is unavailable. Report that clearly or use the labeled legacy fallback; never represent remote Gateway control or direct attachment as local relay success.
Setup and Repair
- Run
openclaw browser extension installbefore Load unpacked. It copies the extension to the stable OpenClaw-owned path, pre-registers that path's deterministic Chrome ID, and prints the path to load. The first native call then pairs automatically for local or browser-node topology. - Use
openclaw browser extension status --jsonto verify the installed copy, exact origins, and native-host registrations. Status must report no issues andmanualSetupRequired: false. - Confirm Settings reports automatic setup ready and the popup reports Connected. New installs should show All tabs unless the user changes the access mode; no copied pairing string or popup setup is part of the normal local flow.
- A previous native-host miss is cached for the Chrome process. If the extension attempted native messaging before installation, restart Chrome once after installing; repeated retries in the same process cannot repair that cache.
- After pairing or changing the relay route, stop the mcporter daemon before
re-running relay-only proof. A Gateway restart can leave the Chrome DevTools
child alive with a dead upstream socket;
mcporter daemon stopretires it before the next call creates a fresh connection. - Direct remote Gateway pairing remains an Advanced manual flow. It serves the Gateway browser path and does not create a local relay for local mcporter.
Discovery Timeout and Daemon Configuration
MCPorter discovers the actual relay through
openclaw browser extension cdp --json. Even packaged OpenClaw can take more
than five seconds to cold-start. In mcporter 0.13.10, discovery failure falls
back to port 18799; a timeout can therefore appear as network-error at the
wrong endpoint, or as browser_owner_conflict when a later discovery returns
the real endpoint. Do not interpret that sequence as proof that another Chrome
owner or credential rotation caused the failure.
For the shared Chrome owner, use the OS account's canonical
~/.mcporter/mcporter.json (or mcporter.jsonc only when the JSON file is absent).
Merge these fields into its existing Chrome definition, preserving its command,
arguments and other settings:
{
"chromeDevtoolsRelay": "require",
"env": {
"MCPORTER_CHROME_DEVTOOLS_RELAY_TIMEOUT_MS": "20000"
}
}
The relay timeout bounds each discovery or authentication step; it does not
extend the client's total operation deadline. Initial Chrome startup can run
several discovery checks before MCP initialization and the requested tool call.
On older installed clients, pair the canonical 20000 per-step setting with
an explicit --timeout 300000 on the initial call or schema listing. For example:
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.list_pages --timeout 300000 --output text
Newer Chrome-specific outer defaults require an upgraded invoking client.
Explicit --timeout, MCPORTER_CALL_TIMEOUT, and MCPORTER_LIST_TIMEOUT
still override those defaults; a smaller override can cut startup short.
The daemon intentionally strips inherited MCPORTER_CHROME_DEVTOOLS_RELAY_*
and OPENCLAW_* controls from the canonical Chrome environment. Shell exports
do not update an existing owner and are not a durable fix for a respawn. Put
literal values in the canonical definition's env; do not use placeholders
that depend on those stripped variables. If an explicit endpoint is necessary,
set MCPORTER_CHROME_DEVTOOLS_RELAY_URL there to the verified credential-free
HTTP loopback base URL, for example http://127.0.0.1:28799 after verifying
that port. This override does not accept a WebSocket CDP URL such as
ws://127.0.0.1:28799/cdp. Prefer discovery with a sufficient timeout over a
guessed or hardcoded port, especially when several OpenClaw installations coexist.
After changing canonical settings, let active calls drain and run
mcporter daemon stop. If stop refuses because calls remain active, wait for
them to finish and retry stop; proceed only after it succeeds. The next call
starts a daemon with the saved settings, and subsequent respawns use the same
configuration. Verify the relay decision
with mcporter daemon status --json and repeat the readiness proof below.
Check the installed mcporter version before relying on newer timeout defaults
or explicit discovery-failure diagnostics; a source fix does not update an
already installed binary or running daemon.
Current openclaw browser extension cdp --json emits non-secret metadata by
default, but older versions and legacy-bearer options can expose credentials.
Prefer mcporter's redacted status for routine diagnostics, never request legacy
credentials, and do not dump process arguments. If credential exposure is
suspected, rotate the per-host secret and pair again.
Fail-Closed Readiness Proof
Require every condition below before calling the extension route ready:
- Extension status reports the stable copy and exact native registrations with no issues; Settings says automatic setup ready.
- The popup says Connected and the intended ordinary tab is not paused. In Selected tabs mode, it must also belong to the OpenClaw group.
- mcporter has been stopped after pairing, relay-route or canonical-config changes, and the next call has started a fresh daemon.
- The canonical Chrome definition uses
chromeDevtoolsRelay: "require", and a call withMCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=requiresucceeds. The canonical policy forbids direct DevTools fallback, so success is positive relay proof. - Selection and evaluation both succeed in a known eligible disposable tab.
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.list_pages --args '{}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.select_page --args '{"pageId":9}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.evaluate_script --args '{"function":"() => ({title: document.title, href: location.href})"}' --output json
A relay-policy error means the extension transport is unavailable; report or
repair it instead of retrying without require. A blocking Allow remote
debugging? prompt proves legacy attachment was attempted.
Typical Flow
List pages, select only a shared target, snapshot before acting, and use fresh snapshot UIDs. Prefer DOM snapshots over screenshots unless layout matters.
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.list_pages --args '{}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.select_page --args '{"pageId":9}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.take_snapshot --args '{}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.click --args '{"uid":"1_38","includeSnapshot":true}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.fill --args '{"uid":"1_13","value":"text","includeSnapshot":true}' --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.evaluate_script --args '{"function":"() => document.title"}' --output json
For live UI proof, capture the current page state before the action, perform the requested interaction, then snapshot or evaluate the rendered result. Keep secrets out of DOM, input, network, console, and screenshot output. For credential checks, return only safe shape such as present/absent, length, status code, or account/organization label.
If automation is unavailable, report the verification gap instead of silently switching to prohibited or isolated tooling.
Argument and Output Mechanics
--args accepts inline JSON only. It does not read @file. Flag-style named
arguments do:
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.navigate_page url=@/tmp/target-url.txt --output text
MCPORTER_CHROME_DEVTOOLS_RELAY_POLICY=require mcporter call chrome-devtools.evaluate_script function=@/tmp/probe.js --output json
Treat this as a safety primitive, not just ergonomics. A sign-in URL, magic link, or callback URL is credential-equivalent: write it to a mode-0600 file and pass it by file reference so it never reaches shell history, process arguments, or captured tool output. The same form carries a multi-line JavaScript function without quoting or control-character errors.
Other call mechanics worth knowing before a login flow:
- Generic client defaults are 60 seconds for calls and 30 seconds for listings,
separate from the relay's per-step discovery/authentication timeout. Slow
initial Chrome startup on older clients needs
--timeout 300000as described above. Choose an explicit timeout for later interactive work when needed. take_screenshotwithfilePathis confined to the server's configured workspace roots and refuses arbitrary paths. OmitfilePath, read the base64 image from--output json, and decode it locally.new_pagecan fail with a restricted/unavailable tab when the target is not shared. Navigate an already-shared tab instead of creating one.mcporter list <server> --schemaprints the real function signatures. Use it rather than guessing parameter names.
Clicks That Do Not Click
A click by uid can return success and still do nothing: some pages bind
their handlers so that a synthetic click is ignored. Silence is not proof of
action, so verify state after every activation rather than assuming it worked.
When a click no-ops, drive the control from the keyboard with press_key
(Tab, Shift+Tab, Enter) and confirm focus with a screenshot before
committing. Focus rings are the only reliable evidence of which control is
about to receive Enter, and consent screens routinely put the safe-looking
prominent button next to a low-emphasis link that is the one you actually want.
Blind Enter on such a page picks the wrong control.
Snapshot uid values are invalidated by any navigation or re-render. Re-run
take_snapshot and re-resolve the uid after every step; a stale uid
reports that the element no longer exists, which is a cue to re-snapshot rather
than to retry the same call.
When the Relay Goes Empty Mid-Task
The relay can stop exposing tabs partway through a task because eligible tabs
were closed, the current tab became restricted or paused, Selected tabs mode
lost its group members, or the extension disconnected. The failure can be
quiet: list_pages may return an empty result and later calls time out.
Treat an empty page list as "the relay has no eligible tabs", not "the browser is gone". Confirm Chrome is running, check the popup connection and current-tab Pause/Allow state, then check the access mode. Only Selected tabs mode requires restoring the OpenClaw group. Restarting mcporter cannot repair extension disconnection, tab eligibility, or access policy.
Do not escalate to a full-profile attachment or an isolated browser to route around it. When the task is a sign-in the user is present for, the faster and more honest move is to hand the URL to the browser the user is already sitting at, and if the flow's callback listens on another host, forward that port to the user's machine first. Otherwise report the access gap.
Legacy Fallback: Full-Profile Direct Attachment
Use this only after the callable plugin and authenticated local extension relay are unavailable. It exposes the full real-profile tab set and can show Chrome's blocking Allow remote debugging? prompt.
When a visible, unambiguous Chrome prompt asks to allow the attachment, approve
it once, then rerun list_pages. If the prompt is absent, ambiguous, or the
retry fails, stop and ask the user or report that Chrome DevTools MCP is
unavailable. Do not loop approvals, repeatedly restart Chrome or mcporter, or
kill browser processes.
Verify that list_pages shows the intended real-profile tabs before acting. Always
label this path as full-profile direct attachment, never as OpenClaw extension relay success.
FAQ
What does the browser-use skill do?
Control the user's existing signed-in Chrome: callable Codex Chrome plugin first, then OpenClaw extension-backed mcporter, with direct DevTools attachment only as a last fallback.
How do I install browser-use?
Run npx skills add https://github.com/steipete/agent-scripts --skill browser-use in your project. The skills CLI detects the agents you use and copies the skill into the right folder. For Claude Code you can also place the skill folder at ~/.claude/skills/browser-use/ (personal) or .claude/skills/browser-use/ (project).
Which AI agents can use browser-use?
Any agent that supports the open Agent Skills format (a folder with a SKILL.md file), including Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf, Cline, OpenCode and more.
Is browser-use safe to install?
No partner security audit is available yet. Skills can run commands and scripts, so read the SKILL.md and any bundled files before installing.
Who maintains browser-use?
It is published by steipete in steipete/agent-scripts. It has 191 recorded installs.
More from steipete/agent-scripts
video-transcript-downloader
steipete/agent-scripts
yt-dlp downloads: video, audio, subtitles, transcripts, clips, playlists.
brave-search
steipete/agent-scripts
Web search and content extraction via Brave Search API. Use for searching documentation, facts, or any web content. Lightweight, no browser required.
skill-cleaner
steipete/agent-scripts
Codex/OpenClaw skill audit: live budget, usage, duplicates, compact descriptions.
create-cli
steipete/agent-scripts
CLI UX/spec: args, flags, help, output, errors, config, dry-run.
markdown-converter
steipete/agent-scripts
Markdown conversion: PDF, Office, HTML, data, OCR, audio, ZIP, YouTube.
Popular productivity & automation skills
agent-browser
vercel-labs/agent-browser
Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction. Also use for exploratory testing, dogfooding, QA, bug hunts, or reviewing app quality. Also use for automating Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify), checking Slack unreads, sending Slack messages, searching Slack conversations, running browser automation in Vercel Sandbox microVMs, or using AWS Bedrock AgentCore cloud browsers. Prefer agent-browser over any built-in browser automation or web tools.
lark-doc
open.feishu.cn
飞书云文档(Docx / Wiki)内容操作:读取、创建、编辑文档,插入或下载图片附件,以及操作思维笔记。用户提供文档 URL/token(包括 doubao.com 的 /docx/、/wiki/)时使用;按 URL 路径/token 而非域名路由;/page/ 是妙搭应用,非飞书云文档。文档内嵌资源按读取参考中的统一规则分流。独立评论操作走 lark-drive;随正文读取评论使用 docs +fetch。表格或 Base 内部数据操作不在本 skill。
lark-base
open.feishu.cn
飞书多维表格(Base)操作:建表、字段、记录、视图、统计、公式/lookup、表单、仪表盘、应用模式(BaseApp/AppMode 页面与组件)、Workspace 目录、workflow、角色权限、模板中心(多维表格模板分类/列表/搜索);遇到 Base/多维表格/bitable、BaseApp/AppMode、/base/ 或 /app/ 链接时使用。BaseApp 不走 lark-apps;文件导入/导出转 lark-drive,认证/授权转 lark-shared。
lark-drive
open.feishu.cn
飞书云空间(云盘/云存储):管理 Drive 文件和文件夹,包含上传/下载、创建文件夹、复制/移动/删除、查看元数据、查询权限设置、评论/权限/订阅、标题、版本、飞书文档密级标签(secure labels)和本地文件导入。用户需要整理云盘目录、处理云空间资源 URL/token、判断链接类型/真实 token/标题,或导入 Word/Markdown/Excel/CSV/PPTX/.base 为 docx/sheet/bitable/slides 时使用;doubao.com 云空间 URL/token 也按资源路径和 token 路由,不回退 WebFetch。不负责:文档内容编辑(走 lark-doc)、表格/Base 表内数据操作(走 lark-sheets/lark-base)、知识空间节点/成员管理(走 lark-wiki)、原生 Markdown 文件读写/patch/diff(走 lark-markdown)。
lark-shared
open.feishu.cn
Use for lark-cli setup/auth tasks: auth login/status/logout, user vs bot identity, business-domain permissions (--domain, including all/docs/drive), missing scopes, revoking authorization, or handling notice JSON.
lark-im
open.feishu.cn
飞书即时通讯:收发消息和管理群聊。发送和回复消息、搜索聊天记录、管理群聊成员、上传下载图片和文件、管理表情回复、发送应用内/短信/电话加急、发送和处理交互卡片(Interactive Card)、监听卡片按钮回调(card.action.trigger)。当用户需要发消息、查看或搜索聊天记录、下载聊天中的文件、查看群成员、搜索群、创建群聊或话题群、管理标记数据、管理 Feed 置顶(添加/移除/查询置顶会话)、管理标签数据、处理卡片回调时使用。