LLM Built-in Tools

During reasoning, the AI can call a set of built-in tools to extend its capabilities. These tools are registered by the backend and require no additional configuration from the host application. The tools fall into two categories: sub-agent management (create, query, stop sub-agents) and user interaction (ask the user questions to gather decisions).

Tool Overview

Tool	Purpose	Interrupts Turn
subAgent	Create a sub-agent session for complex, multi-step tasks	Depends on mode
listSubAgent	List all active sub-agents in the current session tree	No
getSubAgentMessage	Get the latest message from a specific sub-agent	No
stopSubAgent	Stop a specific sub-agent and all its descendant sessions	No
sendMessageToSubAgent	Send an additional message to an async sub-agent	No
askUser	Ask the user 1-4 multiple-choice questions and wait for answers	Yes
todoWrite	Manage task lists (create, update todo items)	No
createGoal	Create an autonomous goal, automatically driving subsequent turns until completion	No
completeGoal	Mark the goal as completed	No
cancelGoal	Cancel the goal	No
createLoop	Create a scheduled loop task that runs automatically at fixed intervals	No
cancelLoop	Cancel a loop task	No
completeLoop	Mark a loop task as completed	No
listLoops	List all loop tasks in the current session	No
pauseLoop	Pause a loop task	No
resumeLoop	Resume a paused loop task	No
webSearch	Search the web for up-to-date information (when configured)	No
webFetch	Fetch a URL and extract content using AI (when configured)	No

💡 Additionally, the memory system provides tools saveMemory, updateMemory, deleteMemory, listMemories, and searchMemory. See Memory System for details.

---

Sub-Agent Tools

Sub-agents are RTC Agent’s task decomposition mechanism. The main agent can break complex tasks into sub-tasks, each running in its own session with an independent context.

Session Hierarchy

Sub-agents form a tree hierarchy through parent/root session IDs:

Field	Description
parent_server_session_id	The server-side UUID of the direct parent session
root_server_session_id	The server-side UUID of the root session (inherited from the parent if the current session is already a sub-session)

---

subAgent — Create a Sub-Agent

Create a new sub-agent session to handle a complex, multi-step task. The sub-agent starts with a blank context, so you need to provide complete background information — like briefing a smart colleague who just walked into the room.

Parameters

Parameter	Type	Required	Description
title	string	Yes	Short task description (3-5 words), e.g. "OAuth authentication setup", "Login button fix"
instruction	string	Yes	Task instruction for the sub-agent. Must include all necessary context since the sub-agent starts with a blank context
mode	string	No	Execution mode: "async" (default) or "sync"

Execution Modes

Mode	Behavior	Use Case
async (default)	Returns the sub-session ID immediately. The parent session continues running. When the sub-agent completes, a notification message is pushed to the parent session	The parent can continue working without waiting for the result
sync	The parent session pauses (via StatefulInterrupt) and waits for the sub-agent to complete, then returns the result	The parent needs the sub-agent’s result before continuing

Usage Tips

• Be specific — include file paths, relevant context, and approaches already ruled out
• Don’t write vague instructions like “handle this task” — the sub-agent doesn’t have the current conversation context
• Keep titles short but descriptive for easy identification in lists

---

listSubAgent — List Sub-Agents

List all active sub-agent sessions in the current session tree. Use this to check the status of spawned sub-agents.

Parameters

None.

Return Value

Returns a JSON array, each element containing:

Field	Type	Description
sub_session_id	string	Server-side UUID of the sub-session
title	string	Sub-agent title
status	string	Session status (only active sessions are returned)
created_at	string	Creation time (RFC 3339 format)
updated_at	string	Update time (RFC 3339 format)

---

getSubAgentMessage — Get Sub-Agent Message

Get the latest message from a specific sub-agent session. Use this to check a sub-agent’s latest output or progress.

Parameters

Parameter	Type	Required	Description
sub_session_id	string	Yes	Server-side UUID of the target sub-session. Use listSubAgent to get available session IDs

Return Value

Returns a JSON object:

Field	Type	Description
sub_session_id	string	Sub-session UUID
sub_session_title	string	Sub-session title
sub_session_status	string	Sub-session status
message_id	string / null	Latest message ID (null if no messages)
role	string / null	Message role (null if no messages)
content	string	Message text content
created_at	string / null	Message creation time (null if no messages)

Permission Check

• The target session must be a descendant of the current session tree (root_server_session_id must match)
• Cannot query the current session itself

---

stopSubAgent — Stop a Sub-Agent

Stop a specific sub-agent session and all its descendant sessions. Uses a bottom-up stopping order (leaf nodes first, then parents).

Parameters

Parameter	Type	Required	Description
sub_session_id	string	Yes	Server-side UUID of the target sub-session. Use listSubAgent to get available session IDs

Return Value

Returns a JSON object:

Field	Type	Description
stopped_sessions	array	List of stopped sessions, each with sub_session_id and title
total_stopped	int	Total number of stopped sessions

Stopping Flow

---

sendMessageToSubAgent — Send Message to Sub-Agent

Send an additional message to an async sub-agent. Used to provide extra instructions, supply additional context, or respond to the sub-agent’s questions during its execution.

Parameters

Parameter	Type	Required	Description
sub_session_id	string	Yes	The server UUID of the target async sub-agent session
message	string	Yes	The message content to send to the sub-agent

Use Cases

• Provide additional instructions during sub-agent execution
• Respond to questions from an async sub-agent
• Redirect the sub-agent’s focus

---

askUser — Ask the User

The AI asks the user 1-4 multiple-choice questions to gather preferences, clarify ambiguity, understand requirements, or get decisions on implementation choices. Users can always pick “Other” to provide free-form text.

Parameters

Parameter	Type	Required	Description
questions	array	Yes	1-4 question objects

Each question object contains:

Field	Type	Required	Description
question	string	Yes	The complete question text, ending with a question mark
header	string	Yes	Short label (max 12 characters) displayed as a chip/tag next to the question
options	array	Yes	2-4 options
multiSelect	boolean	No	Set to true to allow selecting multiple options. Default: false

Each option contains:

Field	Type	Required	Description
label	string	Yes	Option display text (1-5 words)
description	string	Yes	Explanation of what this option means or what will happen if chosen
preview	string	No	Preview content rendered when this option is focused (supports multi-line text, only for single-select questions)

Interaction Flow

Usage Example

{
  "questions": [
    {
      "question": "Which date formatting library should we use?",
      "header": "Date Library",
      "options": [
        {
          "label": "date-fns",
          "description": "Lightweight, tree-shakable, import only what you need"
        },
        {
          "label": "dayjs",
          "description": "Moment.js compatible API, only 2KB"
        }
      ]
    }
  ]
}

Return Format

After the user answers, the AI receives text in this format:

User has answered your questions: "Which date formatting library should we use?"="date-fns". You can now continue with the user's answers in mind.

---

todoWrite — Task List Management

Manage the AI’s task list (Todo List). Uses a full replacement mode — each call passes the complete todo list, replacing the previous content entirely. Updates are published as events to notify the frontend for real-time display.

Parameters

Parameter	Type	Required	Description
todos	array	No	Todo item list (complete replacement)

Each todo item contains:

Field	Type	Required	Description
content	string	Yes	Task description (imperative)
status	"pending" / "in_progress" / "completed"	Yes	Task status
active_form	string	No	Short description for status bar display when in progress

Usage Tips

• Keep tasks at a moderate granularity — each task should be a verifiable milestone
• Update status promptly — mark completed tasks as completed, in-progress ones as in_progress
• Remove unneeded tasks from the list directly instead of marking them as cancelled

---

Goal Tools — Autonomous Goal Driving

The Goal system allows AI to create autonomous goals and automatically advance them across turns until completion. Each Goal has its own state lifecycle and turn-boundary checkpoint.

Goal State

State	Description
active	Goal is active; progress checked each turn
completed	Goal has been completed
cancelled	Goal was cancelled
exhausted	Reached max turn limit; automatically terminated

createGoal — Create a Goal

Parameter	Type	Required	Description
condition	string	Yes	Description of the goal completion condition
max_turns	int	No	Maximum turn limit (default 50)

Execution mechanism: After each Turn completes, the system automatically checks active Goals and increments the completed_turns counter. If the condition is met, the Goal is marked as completed; if max_turns is reached, it is marked as exhausted.

completeGoal / cancelGoal

Parameter	Type	Required	Description
goal_id	string	Yes	Goal ID

---

Loop Tools — Scheduled Recurring Tasks

The Loop system allows AI to create scheduled recurring tasks. Built on the asynq background task queue, it supports pause, resume, and automatic expiry.

Loop State

State	Description
active	Loop is active; triggers automatically at intervals
paused	Loop is paused; no new turns triggered
completed	Loop has completed
cancelled	Loop was cancelled
exhausted	Reached max turns or expired

createLoop — Create a Loop

Parameter	Type	Required	Description
prompt	string	Yes	Prompt to execute on each cycle
interval_seconds	int	Yes	Execution interval (seconds)
max_turns	int	No	Maximum executions (default 10)

Other Loop Operations

Tool	Parameter	Description
cancelLoop	loop_id	Cancel the loop
completeLoop	reason	Mark the loop as completed (with completion reason)
listLoops	None	List all loops in the current session
pauseLoop	loop_id	Pause the loop
resumeLoop	loop_id	Resume a paused loop

💡 Loop vs Goal: A Goal is “work until the condition is met, then stop”; a Loop is “execute every N seconds”. They can be combined — create a Goal inside a Loop to have AI periodically check whether a condition is satisfied.

---

webSearch — Web Search

Search the internet for up-to-date information. Supports multiple search engines (DuckDuckGo, Bing, SearXNG, Tavily) with automatic failover. This tool is only registered when the server has the web_search module configured.

Parameters

Parameter	Type	Required	Description
query	string	Yes	The search query. Should be specific and clear. Examples: "latest Go programming news", "Python asyncio tutorial"
max_results	integer	No	Maximum number of results to return. Default: 10, max: 30. Use fewer for quick answers, more for comprehensive research
time_range	string	No	Limit results to a time range. Options: "day" (last 24h), "week", "month", "year". Leave empty for all time

Return Format

Returns a JSON object containing:

Field	Type	Description
query	string	The original query
result_count	integer	Number of results
results	array	Search result list, each entry contains title, url, description, source

Usage Tips

• Include the current year in your query (e.g. "Go docs 2026") to get the most up-to-date results
• Results include titles, URLs, and snippets — analyze them to find the most relevant URLs
• If snippets are sufficient to answer the question, respond directly; if deeper analysis of a specific URL is needed, use the webFetch tool
• Your response must include a “Sources:” section at the end, listing all referenced source links

---

webFetch — Web Content Fetching

Fetches content from a specified URL and extracts information using an AI model. Converts HTML to Markdown and processes it with a small, fast model. This tool is only registered when the server has the web_fetch module configured.

WARNING: webFetch will FAIL for authenticated or private URLs (e.g. Google Docs, Confluence, Jira, private GitHub repos). For authenticated access, use a specialized MCP tool instead.

Parameters

Parameter	Type	Required	Description
url	string	Yes	The URL to fetch. Must be a fully-formed HTTP or HTTPS address
prompt	string	Yes	Extraction prompt describing what information you want to extract from the page

How It Works

Feature	Description
Cache	Self-cleaning 30-minute cache for faster repeated access to the same URL
Redirects	When a URL redirects to a different host, the redirect URL is returned for a follow-up request
Content limits	Max URL length: 2,000 chars; max content size: 10 MB
Timeout	Fetch timeout: 60 seconds
LLM extraction threshold	LLM extraction is triggered when content exceeds 50,000 bytes (configurable)

---

Next Steps

• Command System — Learn about the slash commands available to users
• Skill System — Learn how the host registers custom functions to extend AI capabilities
• Session Management — Learn about session hierarchy and context management
• Remote Tool Calling — Learn about the remote calling mechanism behind tool invocations