browserpaw
/browserpaw
Steruje aktywnym Chrome użytkownika przez wtyczkę BrowserPaw Hermes: przeglądanie, formularze, klikanie, wyciąganie DOM i tekstu, scrapowanie, zrzuty ekranu i wieloetapowe zadania z numerowaniem elementów DOM.
- Autor
- goldenloaf24h ↗
- Repozytorium
- browserpaw ↗
- Licencja
- AGPL-3.0
- Ostatnia zmiana repozytorium
- 19.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 13
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 19.10.2026, bez odwołań do wycofanych modeli i API.
AGPL-3.0, copyleft: przeróbki udostępniane innym muszą mieć tę samą licencję.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: config/doctor.mjs, config/repair.bat, config/repair.ps1
Co robi i kiedy
Steruje aktywnym Chrome użytkownika przez wtyczkę BrowserPaw Hermes: przeglądanie, formularze, klikanie, wyciąganie DOM i tekstu, scrapowanie, zrzuty ekranu i wieloetapowe zadania z numerowaniem elementów DOM.
Oryginalny opis (EN)
Control and automate the user's active Chrome browser via the BrowserPaw Hermes plugin. Use when asked to browse the web, open URLs, search on Google/Bing/Baidu, fill out forms, click buttons, inspect or extract DOM/text/markdown, scrape web pages, take screenshots, or automate complex multi-step browser tasks. Operates natively inside Chrome with 1-based numeric DOM indexing, autonomous on-page micro-loops (browserpaw_act_toward_goal), and background tab isolation.
Jak działa
- 5-Step Quickstart
- 1. Execution Hierarchy & 3-Tier Routing Ladder
- 2. Tool Selection by Intent (16 Curated Tools)
- 3. On-Page Autonomy: browserpawacttowardgoal
- 4. Operational Invariants
- 5. Failure Recovery Protocols
- 6. References (Consult On Demand)
Podgląd SKILL.md
---
name: browserpaw
description: Control and automate the user's active Chrome browser via the BrowserPaw Hermes plugin. Use when asked to browse the web, open URLs, search on Google/Bing/Baidu, fill out forms, click buttons, inspect or extract DOM/text/markdown, scrape web pages, take screenshots, or automate complex multi-step browser tasks. Operates natively inside Chrome with 1-based numeric DOM indexing, autonomous on-page micro-loops (browserpaw_act_toward_goal), and background tab isolation.
---
# BrowserPaw Browser Control Skill (Hermes Plugin)
Operates directly inside the user's active Chrome session via native CDP (`isTrusted: true`), preserving cookies, logins, and extensions.
> **Tool Availability Notice**: Standard BrowserPaw MCP exposes the full 50 toolset (under full profile; stdio defaults to core). The optional Hermes plugin exposes 16 curated wrapper tools (`browserpaw_*`). Other tools are not available in the Hermes plugin; use standard MCP.
---
## 5-Step Quickstart
1. **Discover Tabs**: Call `browserpaw_get_windows_and_tabs` to check existing tabs before opening duplicates.
2. **Open / Focus Target**: Call `browserpaw_navigate { url, background: true }` (or `browserpaw_switch_tab { tabId }`). Use `background: true` to avoid stealing user focus.
3. **Perceive or Automate**:
- For goal-directed multi-action chains: Call `browserpaw_act_toward_goal { goal: "..." }` when autonomous loop is configured.
- For explicit inspection: Call `browserpaw_read_dom { format: "compact" }` for 1-based numeric indices or `browserpaw_get_markdown` for clean article text.
4. **Interact Precisely**: Call `browserpaw_interact_index { index, action: "click" }` or `browserpaw_fill_index { index, text, pressEnter: true }`. Never guess CSS selectors when an index is known.
5. **Verify & Clean Up**: Inspect return payload or delta. Close temporary task tabs via `browserpaw_close_tabs { tabIds: [...] }`.
---
## 1. Execution Hierarchy & 3-Tier Routing Ladder
Dual-Brain architecture: Agent acts as **System 2 (Macro Planner)** for multi-page orchestration; local server acts as **System 1 (`browserpaw_act_toward_goal`)** executing on-page perceive-decide-act loops.
| Tier | Layer | Primary Tools | Operational Purpose |
| :--------: | :--------------------------- | :------------------------------------------------------------------------------------------------------------- | :--------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Tier 1** | **Semantic Micro-Loop** | `browserpaw_act_toward_goal` | On-page exploratory action goals: searches, exploratory navigation, and multi-action flows when Jev is enabled. |
| **Tier 0** | **Deterministic Primitives** | `browserpaw_get_markdown` / `browserpaw_batch_actions` / `browserpaw_interact_index` / `browserpaw_fill_index` | Deterministic 2+ step form pipelines (`browserpaw_batch_actions`), direct text extraction (`browserpaw_get_markdown`), or candidate takeover on escalation. |
| **Tier 2** | **Visual & Inspection** | `browserpaw_screenshot` / `browserpaw_inspect_media` | Visual inspection, canvas/layout verification, and media extraction. (For raw CDP, in-page JavaScript execution, or synthetic mouse emulation, use standard BrowserPaw MCP). |
---
## 2. Tool Selection by Intent (16 Curated Tools)
| Intent | Tool | Strategy & Key Arguments |
| :----------------------------- | :------------------------------------------------------------------------------------ | :------------------------------------------------------------------------------------------------------------------------------------ |
| **Fuzzy on-page action goals** | `browserpaw_act_toward_goal` | Interactive multi-action on-page goals when Jev is enabled: `{ goal: "Search for laptop and filter by brand Lenovo", maxSteps: 10 }`. |
| Open / back / forward | `browserpaw_navigate` | `{ url }` (`"back"` / `"forward"` for history), `background: true` (default, prevents focus theft). |
| Read article / docs | `browserpaw_get_markdown` | Clean structured text, 80%+ cheaper than full DOM; `fit: true` targets main article body. |
| Find button / text | `browserpaw_grep` | Element lookup without full DOM dump; `autoScroll: true` probes virtual feeds. Returned index feeds `browserpaw_interact_index`. |
| Scroll until element found | `browserpaw_scroll_until_found` | `{ query, maxSteps: 10, stepPx: 800 }`. Client-side scroll, settles virtual DOM, centers element, returns live index. |
| Perceive page structure | `browserpaw_read_dom` | Auto-isolates active modals; `format: "compact"` (default) prunes DOM overhead; `dismissOverlays: true` dismisses modal overlays. |
| Single input / search | `browserpaw_fill_index` | `{ index, text, clear: true, pressEnter: true }` fills and submits in 1 turn. Preserves linebreaks in rich editors. |
| Click / hover element | `browserpaw_interact_index` | `{ index, action: "click" }`. `pierceOverlay: true` (default) clicks beneath translucent masks. |
| 2+ predictable steps | `browserpaw_batch_actions` | Atomic pipeline (fill + click + wait) in 1 RTT. Default for login and search workflows. |
| Scroll page / container | `browserpaw_smart_scroll` | Auto-detects scrollable container; reports remaining scroll distance. |
| Inspect media / audio / video | `browserpaw_inspect_media` | Inspect page images, audio, video elements, and source streams. |
| Page / element screenshot | `browserpaw_screenshot` | Capture visual state for layout inspection or verification. |
| Tab & window hygiene | `browserpaw_get_windows_and_tabs` / `browserpaw_switch_tab` / `browserpaw_close_tabs` | Query existing tabs before opening duplicates. Closing requires explicit `tabIds`. |
| Tool docs & profile query | `browserpaw_tool_docs` | Inspect schemas and documentation for available BrowserPaw tools. |
---
## 3. On-Page Autonomy: `browserpaw_act_toward_goal`
Execute interactive on-page tasks in a single turn:
```json
{
"goal": "Type 'wireless keyboard' into the search bar and submit",
"tabId": 101,
"maxSteps": 10,
"pauseBeforeKeywords": ["Post", "Pay", "Submit"]
}
```
### Status Handling
- `done`: Goal achieved; continue macro workflow.
- `paused`: Triggered by `pauseBeforeKeywords`; inspect state and commit explicitly.
- `escalate`: Ambiguous target or sensitive boundary. **Target candidates already indexed in `currentElements`; pick target `[n]` and call `browserpaw_interact_index` or `browserpaw_fill_index` directly. Do NOT re-invoke `browserpaw_read_dom`**.
- `stuck`: 3 steps without DOM/URL mutation; switch to atomic primitives (`browserpaw_interact_index`, `browserpaw_fill_index`, `browserpaw_batch_actions`).
- `blocked`: Challenge or authentication barrier detected; prompt user to complete verification in Chrome.
- `max_steps` / `timeout`: Step quota reached; inspect `currentElements` and proceed.
---
## 4. Operational Invariants
0. **Unified Action Routing Ladder**:
- **Deterministic 2+ step actions**: Use `browserpaw_batch_actions` directly to execute the atomic pipeline in a single round-trip.
- **Fuzzy exploratory goals**: Dispatch `browserpaw_act_toward_goal` when on-page semantic exploration is needed.
- **Escalation & Takeover**: When `browserpaw_act_toward_goal` returns `escalate` or `paused`, immediately inspect attached `currentElements` candidates and take over with `browserpaw_interact_index` or `browserpaw_fill_index` without re-reading the DOM.
1. **Avoid Step-by-Step Polling**: Avoid manual `browserpaw_read_dom` -> single-click round-trip loops when a batch pipeline or autonomous goal can express the flow.
2. **Numeric 1-Based Indices**: Target elements strictly by `[n]` index from `browserpaw_read_dom`, `browserpaw_grep`, or `currentElements`. Do not guess CSS selectors.
3. **Zero-RTT Escalation Takeover**: When `browserpaw_act_toward_goal` escalates or pauses, consume attached `currentElements` immediately.
4. **Background Tab Discipline**: Open tabs with `background: true` to avoid stealing user OS focus.
5. **Destructive Close Guard**: `browserpaw_close_tabs` requires explicit `tabIds` to protect user tabs.
6. **Single-Turn Submission**: Use `pressEnter: true` on `browserpaw_fill_index` for search boxes and single-field forms.
7. **Overlay Handling**: Pass `pierceOverlay: true` on `browserpaw_interact_index`, or set `dismissOverlays: true` in `browserpaw_read_dom`.
---
## 5. Failure Recovery Protocols
- **Virtualized Feeds / Infinite Scroll**: Invoke `browserpaw_scroll_until_found { query, maxSteps: 10 }` or `browserpaw_grep { query, autoScroll: true }` instead of manual multi-turn scrolling.
- **Modal Occlusion**: `browserpaw_read_dom` auto-isolates modals. If collapsed unexpectedly, pass `isolateModal: false`.
- **Overlay Interference**: Pass `pierceOverlay: true` on `browserpaw_interact_index` or `dismissOverlays: true` on `browserpaw_read_dom`.
- **High-RTT Form Chains**: Consolidate into `browserpaw_batch_actions` with `waitForSettle: true`.
- **Anti-Bot / 2FA / CAPTCHA**: Prompt human user to complete verification in the browser window, then resume.
- **Standard MCP Full Capabilities**: For raw CDP access, in-page JavaScript evaluation, file upload automation, or OS-level mouse control, configure standard BrowserPaw MCP (50 tools).
---
## 6. References (Consult On Demand)
> **Notice**: Reference documents describe the full capabilities of standard BrowserPaw MCP. The optional Hermes plugin is restricted to the 16 curated tools listed above.
- `references/dual-brain-jev.md` — Micro-loop architecture, heuristic fallback, and takeover protocols.
- `references/tool-cheatsheet.md` — Complete contracts and parameter tables for all 50 registered tools.
- `references/batch-pipeline.md` — Multi-action batch grammar and assertions.
- `references/visual-fallback.md` — Screenshot calibration and visual inspection.
- `config/TROUBLESHOOTING.md` — Port conflicts, token auth, and native host repair.
Pliki w paczce (13)
- SKILL.md 12 KB
- config/TROUBLESHOOTING.md 8 KB
- config/TROUBLESHOOTING.zh-CN.md 8 KB
- config/doctor.mjs 15 KB
- config/mcp-config.json 3 KB
- config/repair.bat 1 KB
- config/repair.ps1 1 KB
- recipes/README.md 2 KB
- recipes/template.md 1 KB
- references/batch-pipeline.md 6 KB
- references/dual-brain-jev.md 9 KB
- references/tool-cheatsheet.md 27 KB
- references/visual-fallback.md 4 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · goldenloaf24h/browserpaw | 100 | AGPL-3.0 |
| goldenloaf24h/browserpaw | 100 | AGPL-3.0 |
| goldenloaf24h/browserpaw | 90 | AGPL-3.0 |
Migawka bazy z 13.11.2026 · wersja paczki 07612e7f46b3 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 95browser-automation
Steruje przeglądarką na podstawie zrzutów ekranu, bez przejmowania myszy i klawiatury: otwiera strony, wypełnia formularze, zbiera dane. - 100browser
Udostępnia minimalne narzędzia Chrome DevTools Protocol do automatyzacji i scrapowania: start Chrome, nawigacja, wykonywanie JavaScript, zrzuty ekranu i interaktywny wybór elementów DOM. - 100agent-browser
Automatyzuje przeglądarkę przez CLI agent-browser: nawigacja po stronach, wypełnianie formularzy, klikanie, zrzuty ekranu, wyciąganie danych i testy aplikacji. - 100chrome-use
Obsługuje strony w Chrome użytkownika na jego komputerze przez narzędzie wiersza poleceń chrome-use: otwiera strony, czyta je jako tekst, klika, wypełnia formularze, wgrywa pliki i robi zrzuty ekranu. - 100open-browser-agent
Automatyzuje przeglądarkę przez MCP lub CLI open-browser-agent: przegląda strony, klika, wypełnia formularze i zbiera dane z aplikacji webowych. - 90agent-browser
Automatyzuje zadania w przeglądarce przez CLI dla agentów AI, w tym logowanie do stron, wypełnianie formularzy, zrzuty ekranu, scrapowanie danych i testowanie aplikacji.