Built in Tools EN

🌐 Language / 语言: 中文 · English

Built-in Tools Reference

AI Helper ships with 40+ configurable tools and can be dynamically extended via the MCP protocol. This page lists all tools by category and their purposes for quick reference.

← Back to Wiki Home · Related: Core Features · Agent Service

---

Tool Category Overview

Category	Count	Description
Content Extraction	6	Read page content, structured data, and interactive elements
Page Interaction	5	Click, scroll, wait, dialog handling
Forms & Input	4	Form filling, keyboard input, file upload, dropdown selection
Tab Management	2	Open / switch / close tabs
Bookmarks & History	1	Search browser bookmarks and history
Storage Management	2	Cookies, localStorage management
Network Requests	1	Make HTTP requests
Media & Output	5	Screenshots, clipboard, QR codes, downloads, notifications
Debugging & Development	2	Page appearance tweaks, CDP debugging
AI Collaboration	7	Clarification, task planning, prototypes, retrieval, dispatch, etc.
Agent Service	5	Files, recycle bin, commands, search, Skills
Long-Term Memory	1	Cross-session memory storage and retrieval
Knowledge Base RAG	3	Knowledge-base retrieval and ingestion
MCP Tools	dynamic	Third-party tools extended via the MCP protocol

---

Content Extraction (6)

Tool	Description
page_content	Get page content (text / html formats, cross-tab extraction supported)
extract_data	Extract structured data (table / links / forms / images / metadata, cross-tab supported)
query_elements	Extract interactive elements as a tree snapshot (recommended first; returns stable ref ids for interact_element / fill_form; open overlays lifted to the top, paginate with page / hasMore; re-query after page changes; supports filterByText / countOnly)
search_in_page	Regex search of page text (highlighting supported)
iframe_content	Get iframe content (same-origin, nested supported)
scroll_collect	Scroll and collect long content (deduplicated aggregation)

---

Page Interaction (5)

Tool	Description
interact_element	Page element interaction (click / hover / type input; ref / text / selector locating, ref preferred)
scroll_to	Scroll to a position / element / text (alignment options supported)
wait_element	Wait for an element to appear / disappear (strict visibility detection)
wait_navigation	Wait for page navigation (load / domcontentloaded / networkidle supported)
handle_dialog	Handle page dialogs (alert / confirm / prompt); preset accept / dismiss before the triggering action

---

Forms & Input (4)

Tool	Description
fill_form	Batch form filling (fields accept ref / selector locating; rich-text editors / contenteditable supported)
keyboard_input	Keyboard input (bypasses React controlled components)
file_upload	File upload (DataTransfer injection)
select_dropdown	Dropdown selection (native select + custom components, ref locating supported)

---

Tab Management (2)

Tool	Description
manage_tab	Tab management (open / switch / close / back-forward / reload)
list_tabs	Get the list of all tabs

---

Bookmarks & History (1)

Tool	Description
search_browser_data	Search browser bookmarks and history

---

Storage Management (2)

Tool	Description
manage_cookies	Cookie management (CRUD + bulk clear by domain, requires confirmation)
manage_storage	localStorage / sessionStorage management (clear requires confirmation)

---

Network Requests (1)

Tool	Description
fetch_url	HTTP requests (timeout, retry, exponential backoff, AbortSignal cancellation propagation)

---

Media & Output (5)

Tool	Description
capture_page	Page screenshots (download / visual analysis / download + analysis modes; viewport-only capture)
clipboard	Clipboard operations (copy / paste / get selected page text)
qrcode	Generate QR codes (QRCode library + Canvas fallback)
download_file	Download files (requires confirmation)
notify	Desktop notifications

---

Debugging & Development (2)

Tool	Description
inject_css	Modify page appearance (inject stylesheet / inline style / highlight text)
debug_page	Advanced page debugging via the Chrome DevTools Protocol (attach → operate → detach); supports native input events, XHR/fetch capture, page-context evaluation, full-page screenshots, device emulation, etc.

---

AI Collaboration (7)

Tool	Description
clarify_question	Show a clarification dialog (recommended options, countdown, audio alert; concurrent requests are queued and shown one at a time)
plan_task	Complex task decomposition planning (parallel / sequential / conditional execution)
preview_ui	UI prototype preview and management (preview / get actions)
search_chats	Search conversation memory (current session / all historical sessions)
dispatch_task	Dispatch subtasks to sub-Agents (parallel dispatch supported)
manage_agent	Manage paired Agents (query status, switch Agents)
exec_log	Extract historical execution logs to analyze success paths and failure lessons

---

Agent Service (5, requires Agent)

Tool	Description
agent_file	File operations (read / write / list / delete / download, path sandbox)
agent_trash	Recycle bin management (list / restore deleted files)
agent_exec	Run terminal commands (black / gray / white-list three-tier security, force / timeout supported)
agent_search	Search files (by file name or content, fd / ripgrep accelerated)
agent_skill	Skill loading and execution (load / run actions)

---

Long-Term Memory (1, requires Agent)

Tool	Description
agent_memory	Unified memory management. store for add/update/delete, recall for keyword search, manage for review/cleanup, distinguished by the action parameter

---

Knowledge Base RAG (3, requires Agent)

Registered dynamically only when RAG is enabled in the options page and the Agent-side RAG dependencies are available.

Tool	Description
knowledge_search	Search relevant content in the knowledge base (specify a collectionId, or search across libraries)
knowledge_ingest	Ingest text / local files / URLs into the knowledge base (metadata supported)
knowledge_list	List available knowledge bases and their document counts

---

MCP Tools (dynamic extension)

After connecting a third-party tool server via the MCP protocol, its tools are auto-registered into the system; the count depends on the connected MCP Servers. The latest MCP tool list is pulled from the Agent and injected before each reasoning round. See Agent Service → MCP Protocol Extension.

---

Sensitive-Tool Safety Confirmation

The following tools show a confirmation dialog before operating (5-minute timeout auto-rejects, gated by the reactConfig.toolConfirmationEnabled master switch; when off, sensitive operations run without asking):

• manage_tab (close action), download_file, manage_cookies, manage_storage (clear action)
• agent_file (delete action), agent_exec

💡 Queuing & when the clock starts: parallel subtasks can raise several confirmations / clarifications at once. Both dialog kinds share a single-modal gate and are displayed one at a time in FIFO order (the visible dialog tells you "+N other waiting"), so nothing can cover another dialog's buttons. The timeout countdown only starts once an item is actually displayed — queued requests burn no time, so you can never be denied without ever seeing the dialog. If the side panel is closed or the dialog cannot be delivered, the operation fails closed immediately instead of waiting out the 5 minutes.

💡 Approval scope lives only as long as what it was granted for: "Allow Once" approves that single command; "Allow for This Task" remembers the approval only inside the current task run — it is revoked both when the run finishes and when the next task starts, so it never carries over to later tasks. Parallel commands are settled one by one per toolCallId, so clicking "Allow Once" on one never approves the others.

Agent Command Execution Three-Tier Security

1. Blacklist (always blocked): rm -rf /, mkfs.*, fork bombs, curl-to-shell pipes, etc.
2. Graylist (requires confirmation): sudo / doas, npm install -g, chmod -R 777, git push --force, rm -rf <relative path>, etc.
3. Whitelist (allowed directly): regular commands

The lists are evaluated segment by segment: chained commands, environment-variable prefixes, wrappers such as nohup / env, interpreter payloads (bash -c, cmd /c, powershell -Command) and the inside of $() are each checked separately, so cd /tmp && sudo ... requires confirmation too. See Agent Service → Tiered command control.

---

Tool Configuration

In the side panel's tool selection panel (shortcut Ctrl+T / Cmd+T), you can filter by category, search, and enable / disable tools. Each assistant (Agent) can also have its own toolset. The tool panel listens in real time for MCP tool registration and unregistration.