菜单
Target a real Google Chrome profile for browser automation through Chrome DevTools MCP. Provides the chrome-profile CLI, profile discovery, live DevTools probing guidance, setup playbooks, and URL-anchor tab selection.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/hoanghd218/ZaloCRM --skill ak-chrome-profile命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Search library/framework documentation via llms.txt (context7.com). Use for API docs, GitHub repository analysis, technical documentation lookup, latest library features.
Interview-driven advisory skill. Analyzes a prompt or URL (GitHub issue, spec, doc), scouts the codebase, interviews the user one question at a time to reframe the problem into exact requirements and goals, then delivers honest advice: what to do, what to avoid, better alternatives, benefits and trade-offs. Use when the user asks for advice, a second opinion, a sanity check, requirement reframing before committing to a direction, or wants an existing plan or design pressure-tested (grilled) one question at a time.
Browser and desktop automation through the agent-browser CLI. Use for long autonomous browsing, compact page snapshots, screenshots, form filling, login/test flows without real Chrome profile state, scraping, exploratory QA, cloud browsers, Electron apps, Slack automation, and bug hunts.
正在显示 SKILL.md
| name | ak:chrome-profile |
| description | Target a real Google Chrome profile for browser automation through Chrome DevTools MCP. Provides the chrome-profile CLI, profile discovery, live DevTools probing guidance, setup playbooks, and URL-anchor tab selection. |
| user-invocable | true |
| when_to_use | Invoke when browser automation needs the user's real Chrome profile, cookies, account, or a deterministic profile target. |
| category | dev-tools |
| keywords | ["chrome","browser","profile","mcp","devtools","automation","cookies"] |
| license | MIT |
| allowed-tools | ["Bash(bash:*)","Bash(chrome-profile:*)","Bash(python3:*)","mcp__chrome-devtools__list_pages","mcp__chrome-devtools__select_page","mcp__chrome-devtools__take_snapshot","mcp__chrome-devtools__evaluate_script"] |
| metadata | {"author":"agentkit","version":"1.0.0","compatibility":"Requires Python 3.9+ and Google Chrome stable. Works on macOS, Linux, and Windows."} |
Profile-aware browser automation path when the agent needs the user's real Chrome state: the right Google account, cookies, workspace, tenant, or logged-in product session.
This is not the default for ordinary browser testing. Use ak:agent-browser when a fresh or tool-managed browser is enough. Use project-native Playwright or ak:web-testing for repeatable CI tests. Use this skill when the user's actual Chrome profile state matters.
chrome-profile <key> <url> opens a URL in a configured Chrome profile.chrome-profile open --json <key> <url> emits machine-readable binding data for agents.Profile 17 directory names.chrome-profile doctor tells the agent and user whether the opened tab will be readable through a browser bridge.#cdp-profile=<key>&cdp-open=<token> so the agent can bind to the exact tab from a flat MCP tab list.When this skill is invoked, the agent must lead the user through setup if anything is missing. Do not just say "configure MCP." Run the checks, explain the failing layer, and give the next concrete command or browser action.
Do not treat chrome-profile doctor as the only source of truth. It is a static setup heuristic. Before telling the user that the browser bridge is unavailable, make one live probe through Chrome DevTools MCP if those tools are exposed by the current agent runtime.
Required readiness:
chrome-profile list resolves the requested key.cdp-open=<token> from the CLI output or JSON payload.If the profile mapping fails, pause and guide chrome-profile setup. If the bridge check fails, run the live probe below before asking the user to change MCP setup, relaunch Chrome, or close active tabs.
If doctor reports chrome_devtools_mcp_auto_connect or chrome_devtools_mcp_runtime_probe_required, treat that as a configured bridge candidate and run the live MCP probe; open does not need --force for that state. Use only when the user explicitly wants a tab opened for themselves without agent read-back, or when a live MCP probe succeeds but this older or mismatched CLI still cannot classify the bridge.
--forceRun this probe before declaring "not connected", "no readable bridge", or similar.
chrome-profile doctor reported ok=false.doctor is not usable and no live Chrome DevTools MCP probe succeeds.Do not ask the user to relaunch Chrome as the first response when Chrome DevTools MCP tools are already available. A first tool call can be the action that triggers the browser's remote-control consent prompt.
Install the local CLI shim from the shipped skill directory. Use the path that matches how the skill was installed:
# AgentKit / Claude Code native-skill layout
bash ~/.claude/skills/chrome-profile/scripts/install.sh
bash .claude/skills/chrome-profile/scripts/install.sh
# Current skill directory
bash scripts/install.sh
# Claude Code plugin layout
bash ~/.claude/plugins/<plugin>/skills/ak-chrome-profile/scripts/install.sh
# AgentKit / Codex native-skill layout, global or project-local
bash ~/.agents/skills/chrome-profile/scripts/install.sh
bash .agents/skills/chrome-profile/scripts/install.sh
If none of those paths exists, find the installed chrome-profile/scripts/install.sh under the runtime skill directory and run that script. On Windows, run the sibling install.cmd.
Then run the guided checks:
chrome-profile doctor
chrome-profile setup
chrome-profile list
setup reads Chrome's Local State, proposes stable keys, and writes mappings to:
$XDG_CONFIG_HOME/chrome-profile/profiles.json
That per-machine config survives skill and kit updates. Use chrome-profile setup --yes for non-interactive bootstrap.
The CLI opens tabs in Chrome, but Chrome DevTools MCP must be able to read those tabs. doctor summarizes static setup. The live MCP probe above confirms whether the current agent runtime can actually reach the browser.
Use this when the runtime can expose Chrome DevTools MCP without a fixed remote-debugging endpoint. This avoids depending on a runtime-specific browser extension.
chrome-profile doctor
Static doctor output may still be conservative. The live Chrome DevTools MCP page-list/read probe is the final reachability check.
Use this for pure CDP workflows, CI-like local setups, or when auto-connect is not desired. This usually requires relaunching Chrome.
open -na "Google Chrome" --args \
--remote-debugging-port=9222 \
--remote-allow-origins=*
.claude/.mcp.json:{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest", "--browserUrl", "http://127.0.0.1:9222"]
}
}
}
chrome-profile doctor
Expected result:
bridge=chrome_devtools_mcp_attached
ok=true
If doctor reports chrome_devtools_mcp_auto_connect or chrome_devtools_mcp_runtime_probe_required, the static check found Chrome DevTools MCP config but cannot prove runtime reachability; run the live Chrome DevTools MCP probe. If it reports cdp_endpoint_without_mcp_config, Chrome is listening but the MCP config is missing or not loaded. If it reports none, run the live probe before deciding that no readable bridge is available.
Decision step: first decide whether this browser action needs a specific real Chrome profile: cookies, account state, workspace, tenant, or isolated profile key. If no real profile state is required, use ak:agent-browser, ak:web-testing, project-native browser tests, or ordinary Chrome DevTools MCP navigation instead. Continue with this workflow only when the profile identity matters.
When profile identity matters, open the target URL in the intended profile. Agents should prefer JSON output because it gives an exact selector:
chrome-profile open --json work "https://github.com/org/repo/pulls"
Then operate through the active MCP bridge:
bind_selector from the JSON output, for example cdp-open=9f3a....cdp-open=<token>.cdp-profile=work as a CLI marker sanity check.If a human-oriented command was used instead of --json, read the printed find: line and bind to that exact cdp-open=<token> marker. Do not choose a tab by "newest matching cdp-profile" unless the page stripped the open marker after you already captured the page ID.
For background work on macOS, add --no-activate or set CHROME_PROFILE_NO_ACTIVATE=1; the CLI opens the profile tab and then returns focus to the previously active app.
For profile-scoped work, do not use an MCP new_page or raw navigate_page tool as the tab-creation/navigation path. It opens or navigates whichever profile/page the bridge currently targets. Always materialize profile-specific tabs with chrome-profile open --json <key> <url>. If a live MCP probe succeeds but an older or mismatched CLI still refuses the bridge, use chrome-profile open --json <key> <url> --force, then immediately bind to the page whose URL contains the returned bind_selector.
chrome-profile open --json work https://example.com prints:
{
"bind_selector": "cdp-open=6d4f8b0a9c1e2d33",
"open_marker": "cdp-open=6d4f8b0a9c1e2d33",
"opened_url": "https://example.com#cdp-profile=work&cdp-open=6d4f8b0a9c1e2d33",
"profile_dir": "Profile 7",
"profile_key": "work",
"profile_label": "w***@example.com",
"profile_marker": "cdp-profile=work"
}
Treat bind_selector as the primary MCP tab selector. profile_marker is a CLI marker sanity check and stale-tab fallback, not proof of the browser account identity.
| Limit | Handling |
|---|---|
SPAs may rewrite location.hash | Capture the page ID immediately after opening, then keep using that page ID. |
| Missing profile key | Run chrome-profile setup or edit the per-machine config. |
| Unresolved key | The Google account is not signed into this Chrome profile on this machine. Sign in once through Chrome UI. |
| No bridge | Run chrome-profile doctor, then the live MCP probe. Follow Option A or B only if both fail. |
| Non-Chrome browsers | The shipped CLI targets Google Chrome stable only. |
Local State; it does not read cookies, passwords, or profile databases.chrome-profile open --json payload unless the user explicitly asks or the local task needs it.references/architecture.md - why profile targeting works with Chrome's single-process profile model.references/mcp-config-recipes.md - detailed bridge setup recipes and troubleshooting.references/troubleshooting.md - common failures and fixes.