一键导入
operate-browser
// Execute browser operations as instructed and report page state. Use when navigating pages, clicking, filling forms, scrolling, checking page state, or performing other browser interactions without making decisions.
// Execute browser operations as instructed and report page state. Use when navigating pages, clicking, filling forms, scrolling, checking page state, or performing other browser interactions without making decisions.
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
コンテキスト圧迫時に次セッションへの引き継ぎファイルを作成する
You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.
Launch Chrome with remote debugging port and connect agent_browser with --auto-connect. Use when you need to reuse the user's existing browser profile (cookies, logins, extensions) for authenticated browsing automation.
プロジェクトの初期化・設定に関するナレッジ
Verify application behavior for a requested scope and report QA results. Use when performing behavioral QA, exploratory E2E checks, test coverage review, or verification reports.
| name | operate-browser |
| description | Execute browser operations as instructed and report page state. Use when navigating pages, clicking, filling forms, scrolling, checking page state, or performing other browser interactions without making decisions. |
| disable-model-invocation | false |
| user-invocable | true |
Execute browser operations exactly as instructed and report the resulting page state.
**Responsibilities**: - Perform browser operations as directed (navigation, clicks, form input, scrolling, etc.) - Report page state accurately after each operation (content, errors, URL, etc.)Out of scope:
Prefer Claude in Chrome (mcp__claude-in-chrome__*) when available. Fall back to playwright-cli (via Bash) otherwise.
Call mcp__claude-in-chrome__tabs_context_mcp at session start. If it succeeds, use Claude in Chrome. If it errors, fall back to playwright-cli.
Run playwright-cli via the Bash tool. Recommended timeout: 15000ms per command.
1. open [url] # Launch browser (URL optional)
2. state-load <file> # Load auth state if available
3. goto <url> # Navigate to target URL
4. snapshot # Get page structure (obtain element refs)
5. click <ref> / fill <ref> <text> / press <key> # Interact
6. snapshot # Verify state after interaction
7. close # Close browser when done
| Command | Purpose |
|---|---|
open [url] | Launch browser. Blank page if URL omitted |
goto <url> | Navigate to URL |
snapshot | Get page structure as YAML with element ref IDs |
screenshot | Save viewport screenshot as PNG |
click <ref> | Click an element |
fill <ref> <text> | Fill text into an input field |
type <text> | Type text into the active element |
press <key> | Press a key (Enter, Tab, Escape, etc.) |
select <ref> <value> | Select a dropdown option |
tab-list | List all tabs |
tab-new [url] | Open a new tab |
tab-select <index> | Switch to a tab |
close | Close the browser |
Snapshots return YAML. Each element has a [ref=eXX] identifier. Use this ref to target elements.
- search [ref=e26]:
- combobox "Search" [active] [ref=e37]
- button "Search" [ref=e40]
→ fill e37 "query" / click e40
Use -s=<name> to isolate multiple browser sessions:
playwright-cli -s=session1 open https://example.com
playwright-cli -s=session1 snapshot
Auth state is managed in a single file: ~/.playwright-cli/auth-state.json. All cookies/storage across sites are consolidated here.
playwright-cli open
playwright-cli state-load ~/.playwright-cli/auth-state.json # Errors harmlessly if file missing
playwright-cli goto <url>
Always run state-load — it fails gracefully when the file does not exist.
If a page redirects to a login screen, follow these steps strictly in this order:
Step 1: Open headed browser
playwright-cli close # Close current browser if open
playwright-cli open --headed <login-url>
playwright-cli state-load ~/.playwright-cli/auth-state.json
Step 2: Ask the user to log in and STOP Report: "Browser is open. Please log in. Let me know when you're done." You MUST stop here. Do NOT close the browser.
Step 3: After user confirms login (resume from here) You MUST save state BEFORE closing. NEVER close before saving.
mkdir -p ~/.playwright-cli
playwright-cli state-save ~/.playwright-cli/auth-state.json
playwright-cli close
Step 4: Continue in headless mode
playwright-cli open
playwright-cli state-load ~/.playwright-cli/auth-state.json
playwright-cli goto <target-url>
Report after each operation: