Tool Calling

Overview

Tool calling lets AI models perform actions beyond text generation. Instead of saying “I don’t know the current weather,” the model can call a web search tool to find the answer. Off Grid includes built-in tools for web search, calculations, date/time, and device information — all running on-device except web search (which requires network).

Available Tools

Web Search

Search the web for current information via Brave Search.Use cases:

“What’s the weather in New York?”
“Latest news about AI”
“Who won the Super Bowl in 2024?”

How it works:

Scrapes Brave Search for top 5 results
Returns titles, snippets, and clickable URLs
URLs open in your system browser when tapped

Requirements:

Network connection (only tool that requires internet)
Empty queries fall back to your last message

Response format:

1. [Title](https://example.com)
   Snippet text...

2. [Title](https://example.com)
   Snippet text...

Web search sends requests to Brave Search. If you want 100% offline operation, disable this tool in settings.

Calculator

Evaluate mathematical expressions safely.Supported operations:

Basic: +, -, *, /, %
Exponentiation: ^ (e.g., 2^8 = 256)
Parentheses: (, )
Decimals: 3.14, 0.5

Examples:

“What’s 15% of 230?”
“Calculate (45 + 17) * 3”
“2^10 equals what?”

How it works:

Recursive descent parser (no eval() — 100% safe)
Supports nested parentheses and operator precedence
Returns the expression and result: (45 + 17) * 3 = 186

Limitations:

Only basic math operations (no trigonometry, logarithms, etc.)
No variables or functions

Date & Time

Get the current date and time with optional timezone support.Use cases:

“What time is it?”
“What’s the date today?”
“What time is it in Tokyo?”

Response format:

Current date and time: Friday, March 3, 2026, 10:42:15 AM Pacific Standard Time
ISO 8601: 2026-03-03T10:42:15.123Z
Unix timestamp: 1772567735

Timezone support:

Uses IANA timezone database (e.g., America/New_York, Asia/Tokyo)
Defaults to device timezone if not specified
Gracefully handles invalid timezones

Device Info

Retrieve device hardware information.Available info types:

Battery: Level and charging status
Memory: Total, used, and available RAM
Storage: Total and free disk space
All: Device model, OS, and all of the above

Examples:

“How much battery do I have?”
“Check available storage”
“What’s my device info?”

Response format (all):

Memory:
  Total: 8.0 GB
  Used: 4.2 GB
  Available: 3.8 GB

Storage:
  Total: 128.0 GB
  Free: 45.3 GB

Battery: 73% (charging)

Device: Samsung Galaxy S23
OS: android 14

Privacy:

All data comes from react-native-device-info (local queries only)
No network requests
No data leaves your device

URL Reader

Fetch and read the content of a web page.Use cases:

“Summarize this article: https://example.com/article”
“What does this page say? https://news.com/story”

How it works:

Fetches the URL via HTTP GET
Strips HTML tags to extract readable text
Truncates to 4000 characters to fit context window

Security:

Blocks private/local URLs (localhost, 192.168.x.x, 10.x.x.x, cloud metadata endpoints)
SSRF protection — Prevents access to internal network resources
15-second timeout — Prevents hanging on slow servers

Limitations:

Only works with public HTTP/HTTPS URLs
JavaScript-heavy sites may not render properly (fetches HTML source)
Some sites block mobile user agents

URL Reader requires network access. Content is fetched in real-time and may be subject to the website’s terms of service.

Which Models Support Tool Calling?

Not all models support tool calling. Off Grid automatically detects tool calling capability at model load time by inspecting the model’s jinja chat template.

Supported Models

Qwen 3 (all sizes) ✅
Gemma 3 ✅
Llama 3.2 ✅
Phi-4 ✅
Command-R ✅
Mistral (most recent versions) ✅

Unsupported Models

Older Llama models (pre-3.2) ❌
Small instruction models without tool schemas ❌
Base models (non-instruct) ❌

If a model doesn’t support tool calling, the Tools button in the UI is disabled.

How to Use

1. Load a Tool-Compatible Model

Download a model that supports tool calling (e.g., Qwen 3 3B)
Load it in a conversation
Check if the Tools button appears in the chat input (next to the microphone)
If the button is grayed out, the model doesn’t support tools

2. Enable Tools

Tap the Tools button to toggle tool calling on/off for the current conversation:

On (highlighted) — Model can call tools
Off (gray) — Model only generates text

3. Configure Which Tools Are Available

Go to Settings → Model Settings → Text Generation → Enabled Tools
Toggle individual tools on/off:
- ✅ Web Search
- ✅ Calculator (enabled by default)
- ✅ Date & Time (enabled by default)
- ✅ Device Info
- ✅ URL Reader

Only enabled tools are sent to the model. Disabled tools are not available for calling.

Tool configuration is global — applies to all models and conversations. You can toggle tools on/off per conversation via the Tools button.

4. Ask Questions That Require Tools

Just ask naturally — the model decides when to use tools: Examples:

“What’s the weather in Tokyo?” → Web Search
“Calculate 15% of 230” → Calculator
“What time is it in London?” → Date & Time
“How much RAM do I have?” → Device Info
“Summarize this page: https://…” → URL Reader

You’ll see:

Tool call indicator — Shows which tool(s) the model is calling
Tool results — The tool’s output (e.g., search results, calculation answer)
Final response — The model synthesizes tool results into a natural answer

Tool Loop Behavior

To prevent runaway loops and excessive API calls, Off Grid enforces limits:

Max iterations: 3 per generation
Max total tool calls: 5 across all iterations

If these limits are hit, the model is forced to respond with the information it has.

How the Loop Works

1. User sends message
2. Model generates response
3. If response contains tool calls:
   → Execute tools
   → Inject results back into conversation
   → Model continues generating (iteration 2)
4. Repeat until model stops calling tools or limits are hit
5. Final response is returned to user

Fallback XML Parsing

Some smaller models don’t support structured tool calls but can emit XML tags:

<tool_call>
  <name>calculator</name>
  <arguments>{"expression": "15 * 20"}</arguments>
</tool_call>

Off Grid detects these and executes the tool, allowing even simpler models to use tools.

Examples

Web Search Example

User: “What’s the latest iPhone?” Model (internal): Calls web_search tool with query “latest iPhone 2026” Tool Result:

1. [Apple iPhone 16 Pro](https://apple.com/iphone)
   The new iPhone 16 Pro features...

2. [iPhone 16 Review - TechCrunch](https://techcrunch.com/...)
   Apple's latest flagship...

Model (final response): “The latest iPhone is the iPhone 16 Pro, which features… [Full description based on search results]“

Calculator Example

User: “What’s 23% of 890?” Model (internal): Calls calculator tool with expression “890 * 0.23” Tool Result: 890 * 0.23 = 204.7 Model (final response): “23% of 890 is 204.7.”

Multi-Tool Example

User: “What time is it in Tokyo and how much battery do I have?” Model (internal): Calls both get_current_datetime (timezone: Asia/Tokyo) and get_device_info (info_type: battery) Tool Results:

Date/Time: Current date and time: Saturday, March 4, 2026, 2:42 AM Japan Standard Time...
Device Info: Battery: 73% (charging)

Model (final response): “It’s currently 2:42 AM on Saturday, March 4, 2026 in Tokyo. Your battery is at 73% and charging.”

Technical Details

Tool Definition Schema

Tools are defined in src/services/tools/registry.ts:

export const AVAILABLE_TOOLS: ToolDefinition[] = [
  {
    id: 'calculator',
    name: 'calculator',
    displayName: 'Calculator',
    description: 'Evaluate mathematical expressions',
    icon: 'hash',
    parameters: {
      expression: {
        type: 'string',
        description: 'The mathematical expression to evaluate',
        required: true,
      },
    },
  },
  // ...
];

OpenAI Schema Conversion

Tools are converted to OpenAI’s function calling schema and passed to llama.rn:

{
  "type": "function",
  "function": {
    "name": "calculator",
    "description": "Evaluate mathematical expressions",
    "parameters": {
      "type": "object",
      "properties": {
        "expression": {
          "type": "string",
          "description": "The mathematical expression to evaluate"
        }
      },
      "required": ["expression"]
    }
  }
}

Tool Execution

Tool calls are executed in src/services/tools/handlers.ts:

export async function executeToolCall(call: ToolCall): Promise<ToolResult> {
  const start = Date.now();
  try {
    let content: string;
    switch (call.name) {
      case 'calculator':
        content = handleCalculator(call.arguments.expression);
        break;
      // ...
    }
    return { toolCallId: call.id, name: call.name, content, durationMs: Date.now() - start };
  } catch (error) {
    return { toolCallId: call.id, name: call.name, content: '', error: error.message, durationMs: Date.now() - start };
  }
}

Each tool returns a ToolResult with:

toolCallId — Matches the model’s call ID
name — Tool name
content — Tool output (or empty if error)
error — Error message (if tool failed)
durationMs — Execution time

System Prompt Hint

When tools are enabled, Off Grid injects a hint into the system prompt:

You have access to the following tools. Use them when appropriate by making tool calls:
- calculator: Evaluate mathematical expressions
- get_current_datetime: Get the current date and time
- get_device_info: Get device hardware information

When a user asks about current information, real-time data, or anything you don't know, USE the appropriate tool instead of saying you cannot help.

This guides the model to use tools proactively.

Tips

Getting the Best Results

Use tool-compatible models — Qwen 3, Gemma 3, Llama 3.2 work best
Enable only needed tools — Fewer tools = less confusion for the model
Ask specific questions — “What’s 23% of 890?” vs. “Do some math”
Disable tools when not needed — Tap the Tools button to toggle off
Check network for web search — Web search requires internet, other tools work offline

Troubleshooting

Tools button is grayed out:

Your model doesn’t support tool calling
Try loading a different model (Qwen 3, Gemma 3, Llama 3.2)

Model doesn’t call tools:

Ensure tools are enabled (Tools button highlighted)
Check that the tool is enabled in settings (Settings → Model Settings → Enabled Tools)
Try rephrasing your question more explicitly

Web search returns no results:

Check your network connection
Brave Search may be blocking requests (rare)
Try a different search query

Calculator returns an error:

Ensure the expression only uses supported operators (+, -, *, /, %, ^, (, ))
No variables or functions (e.g., sin(), log()) — use a more capable calculator app for advanced math

Tool loop hits max iterations:

The model is stuck in a loop
Try disabling tools and asking the question again
Or rephrase to be more specific

Privacy

Most tools run 100% on-device:

Calculator — No network, all local
Date & Time — No network, all local
Device Info — No network, all local

Network-dependent tools:

Web Search — Sends queries to Brave Search (HTTPS)
URL Reader — Fetches the specified URL (HTTPS)

If you want 100% offline operation, disable Web Search and URL Reader in settings.

Get Started

Core Features

Guides

Overview

Available Tools

Web Search

Calculator

Date & Time

Device Info

URL Reader

Which Models Support Tool Calling?

Supported Models

Unsupported Models

How to Use

1. Load a Tool-Compatible Model

2. Enable Tools

3. Configure Which Tools Are Available

4. Ask Questions That Require Tools

Tool Loop Behavior

How the Loop Works

Fallback XML Parsing

Examples

Web Search Example

Calculator Example

Multi-Tool Example

Technical Details

Tool Definition Schema

OpenAI Schema Conversion

Tool Execution

System Prompt Hint

Tips

Getting the Best Results

Troubleshooting

Privacy

Build docs developers (and LLMs) love

Get Started

Core Features

Guides

​Overview

​Available Tools

​Web Search

​Calculator

​Date & Time

​Device Info

​URL Reader

​Which Models Support Tool Calling?

​Supported Models

​Unsupported Models

​How to Use

​1. Load a Tool-Compatible Model

​2. Enable Tools

​3. Configure Which Tools Are Available

​4. Ask Questions That Require Tools

​Tool Loop Behavior

​How the Loop Works

​Fallback XML Parsing

​Examples

​Web Search Example

​Calculator Example

​Multi-Tool Example

​Technical Details

​Tool Definition Schema

​OpenAI Schema Conversion

​Tool Execution

​System Prompt Hint

​Tips

​Getting the Best Results

​Troubleshooting

​Privacy

Build docs developers (and LLMs) love

Overview

Available Tools

Web Search

Calculator

Date & Time

Device Info

URL Reader

Which Models Support Tool Calling?

Supported Models

Unsupported Models

How to Use

1. Load a Tool-Compatible Model

2. Enable Tools

3. Configure Which Tools Are Available

4. Ask Questions That Require Tools

Tool Loop Behavior

How the Loop Works

Fallback XML Parsing

Examples

Web Search Example

Calculator Example

Multi-Tool Example

Technical Details

Tool Definition Schema

OpenAI Schema Conversion

Tool Execution

System Prompt Hint

Tips

Getting the Best Results

Troubleshooting

Privacy