Tools & tool servers

The agent acts through tools. AI Labs ships a set of built-in tools grouped by capability, and can connect external tools over MCP or OpenAPI.

Built-in tool groups

Group	Tools	What it does
files	read_file, list_directory, search_files, create_file, display_file, edit_file, multi_edit_file, write_file	Read, write, edit, and search the workspace.
terminal	run_command, send_input, check_task, kill_task	Run shell commands foreground or background, feed input to a running process, poll or kill a background task.
web	web_search, read_url	Search the web and fetch page contents.
browser	browser_navigate, browser_snapshot, browser_click, browser_type, browser_screenshot, browser_evaluate	Drive a real browser — navigate, read the accessibility snapshot, click, type, screenshot, run JS.
memory	update_memory	Write to the managed memory vault.
chats	search_chats	Search past conversations.
skills	view_skill, manage_skill	Open a skill and its companion files; create or edit managed skills.
tasks	update_tasks	Maintain a task list for multi-step work.
automations	list_automations, create_automation, update_automation, toggle_automation, delete_automation	Manage scheduled automations from chat.
images	image_generate	Generate and edit images.
subagents	delegate_task	Spawn a sub-agent.
notifications	notify	Send a notification to a configured target.

Admins can disable whole groups. In global (workspace-less) chats, the filesystem, terminal, browser, automation, and image groups plus manage_skill are off by default.

Browser tools

The browser group needs Chromium. In Docker, use the --target browser image or docker compose --profile browser up (see Docker). Screenshots can be size-constrained; browser control is resilient to slow pages.

Web search providers

web_search can be backed by Brave, DuckDuckGo, Exa, Tavily, Perplexity, SearXNG, Firecrawl, or any chat-completions endpoint. Base URLs (e.g. Perplexity) are configurable. Set the provider and keys in Admin → Web.

Tool servers (MCP & OpenAPI)

Connect external tools without writing glue code:

• MCP — local stdio servers (a command AI Labs launches) or remote MCP endpoints. Local servers can carry their own environment variables. MCP errors surface as readable messages; quoted commands are handled correctly.
• OpenAPI — point at an OpenAPI spec and its operations become callable tools.

Manage servers in Admin → Tool Servers: add, edit, delete, and verify connectivity. Tool-server tool calls are visible in chat and through the Gateway API.

Approval & Auto mode

Every tool call runs under an approval policy:

Policy	Behavior
allow	Runs without asking.
review	Pauses for your approval (approve / reject, mouse or keyboard).

Auto mode lets AI Labs approve safe, expected calls itself and still route anything risky or ambiguous through the same review path — so approval prompts stay consistent whether a session is ask-first or auto. Admins pick the review model and per-tool defaults in Admin → Tools (Tool approval). Rejected tools are handled cleanly and the chat picks up where it left off.