The tool() function creates type-safe tools with Zod schema validation:
import { OpenRouter, tool } from '@openrouter/agent';
import { z } from 'zod';
const weatherTool = tool({
name: 'get_weather',
description: 'Get the current weather for a location',
inputSchema: z.object({
location: z.string().describe('City name, e.g., "San Francisco, CA"'),
}),
outputSchema: z.object({
temperature: z.number(),
conditions: z.string(),
}),
execute: async (params) => {
// params is typed as { location: string }
const weather = await fetchWeather(params.location);
return {
temperature: weather.temp,
conditions: weather.description,
};
},
});
The SDK supports four types of tools, automatically detected from your configuration:
Standard tools with an execute function:
const calculatorTool = tool({
name: 'calculate',
description: 'Perform a mathematical calculation',
inputSchema: z.object({
expression: z.string().describe('Math expression like "2 + 2"'),
}),
outputSchema: z.object({
result: z.number(),
}),
execute: async (params) => {
const result = eval(params.expression); // Use a safer eval in production
return { result };
},
});
Tools that yield progress updates during execution. Add eventSchema to enable generator mode:
const searchTool = tool({
name: 'search_database',
description: 'Search documents with progress updates',
inputSchema: z.object({
query: z.string(),
limit: z.number().default(10),
}),
// eventSchema triggers generator mode
eventSchema: z.object({
progress: z.number().min(0).max(100),
message: z.string(),
}),
outputSchema: z.object({
results: z.array(z.string()),
totalFound: z.number(),
}),
// execute is now an async generator
execute: async function* (params) {
yield { progress: 0, message: 'Starting search...' };
const results = [];
for (let i = 0; i < 5; i++) {
yield { progress: (i + 1) * 20, message: `Searching batch ${i + 1}...` };
results.push(...await searchBatch(params.query, i));
}
// Final yield is the output
yield { progress: 100, message: 'Complete!' };
// Return the final result (or yield it as last value)
return {
results: results.slice(0, params.limit),
totalFound: results.length,
};
},
});
Progress events are streamed to consumers via getToolStream() and getFullResponsesStream().
Tools without automatic execution - you handle the tool calls yourself:
const manualTool = tool({
name: 'send_email',
description: 'Send an email (requires user confirmation)',
inputSchema: z.object({
to: z.string().email(),
subject: z.string(),
body: z.string(),
}),
execute: false, // Manual handling required
});
Use getToolCalls() to retrieve manual tool calls for processing.
HITL tools extend manual-tool semantics with two sync-or-async hooks that let you decide per call whether to respond programmatically or pause for a human:
onToolCalled — fires when the model invokes the tool. Return a value to feed the model directly (like a regular execute), or return null to pause the loop like a manual tool. The caller resumes later by supplying a function_call_output item.
onResponseReceived — optional. Fires on a later turn when an incoming function_call_output matches a prior call of this tool (by callId → function_call.name). It receives the caller-supplied raw result and returns the value sent to the model. Throwing surfaces as a tool error to the model.
An outputSchema is required for HITL tools — it validates both the onToolCalled return value (when non-null) and the value delivered via function_call_output (whether transformed by onResponseReceived or passed through directly).
const approvePaymentTool = tool({
name: 'approve_payment',
description: 'Approve a payment, escalating large amounts to a human',
inputSchema: z.object({
amount: z.number(),
recipient: z.string(),
}),
outputSchema: z.object({
ok: z.boolean(),
reviewedAt: z.number().optional(),
}),
onToolCalled: async (input) => {
// Auto-approve small amounts
if (input.amount < 100) {
return { ok: true };
}
// Escalate to a human — pauses the loop
return null;
},
onResponseReceived: async (raw) => {
// Post-process the caller-supplied result before the model sees it
return { ...(raw as object), reviewedAt: Date.now() };
},
});
When onToolCalled returns null, the conversation state moves to status: 'awaiting_hitl' and the paused call surfaces via getToolCalls() / getPendingToolCalls(). Resume by calling callModel again with a function_call_output item for each paused call in the input.
HITL tools differ from requireApproval: approval gates pause before execution for a yes/no decision, while HITL tools let onToolCalled run arbitrary logic first and only pause when it returns null. Use HITL when the decision is data-driven (e.g., amount thresholds, risk scoring); use requireApproval when you always want explicit human consent. See Tool Approval & State.
Schema Definition
Define what parameters the tool accepts:
const inputSchema = z.object({
// Required parameters
query: z.string().describe('Search query'),
// Optional with default
limit: z.number().default(10).describe('Max results'),
// Optional without default
filter: z.string().optional().describe('Filter expression'),
// Enum values
sortBy: z.enum(['relevance', 'date', 'popularity']).default('relevance'),
// Nested objects
options: z.object({
caseSensitive: z.boolean().default(false),
wholeWord: z.boolean().default(false),
}).optional(),
// Arrays
tags: z.array(z.string()).optional(),
});
Output Schema
Define the structure of results returned to the model:
const outputSchema = z.object({
results: z.array(z.object({
id: z.string(),
title: z.string(),
score: z.number(),
})),
metadata: z.object({
totalCount: z.number(),
searchTimeMs: z.number(),
}),
});
Define progress/status events for generator tools:
const eventSchema = z.object({
stage: z.enum(['initializing', 'processing', 'finalizing']),
progress: z.number(),
currentItem: z.string().optional(),
});
Type Inference
The SDK provides utilities to extract types from tools:
import type { InferToolInput, InferToolOutput, InferToolEvent } from '@openrouter/agent';
// Get the input type
type WeatherInput = InferToolInput<typeof weatherTool>;
// { location: string }
// Get the output type
type WeatherOutput = InferToolOutput<typeof weatherTool>;
// { temperature: number; conditions: string }
// Get event type (generator tools only)
type SearchEvent = InferToolEvent<typeof searchTool>;
// { progress: number; message: string }
const openrouter = new OpenRouter({ apiKey: process.env.OPENROUTER_API_KEY });
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather in Tokyo?',
tools: [weatherTool],
});
// Tools are automatically executed
const text = await result.getText();
// "The weather in Tokyo is 22°C and sunny."
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Search for TypeScript tutorials and calculate 2+2',
tools: [searchTool, calculatorTool],
});
Use as const for full type inference on tool calls:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather?',
tools: [weatherTool, searchTool] as const,
maxToolRounds: 0, // Get tool calls without executing
});
// Tool calls are typed as union of tool inputs
for await (const toolCall of result.getToolCallsStream()) {
if (toolCall.name === 'get_weather') {
// toolCall.arguments is typed as { location: string }
console.log('Weather for:', toolCall.arguments.location);
}
}
Execute Context
Tool execute functions receive a flat context object as
their second argument. It merges TurnContext fields
with a tools map and a setContext() method:
const contextAwareTool = tool({
name: 'context_tool',
inputSchema: z.object({ data: z.string() }),
outputSchema: z.object({ result: z.string() }),
execute: async (params, context) => {
// TurnContext fields are available directly
console.log('Turn:', context.numberOfTurns);
console.log('History:', context.turnRequest?.input);
console.log('Model:', context.turnRequest?.model);
return {
result: `Processed on turn ${context.numberOfTurns}`,
};
},
});
Context Properties
| Property | Type | Description |
|---|
numberOfTurns | number | Current turn number (1-indexed) |
turnRequest | OpenResponsesRequest | undefined | Current request object |
toolCall | OpenResponsesFunctionToolCall | undefined | The tool call being executed |
local | Readonly<TContext> | This tool’s own context (read-only) |
setContext | (partial: Partial<TContext>) => void | Mutate this tool’s context |
shared | Readonly<TShared> | Shared context visible to all tools |
setSharedContext | (partial: Partial<TShared>) => void | Mutate shared context |
Tool Context
Tools can declare a contextSchema to receive typed,
persistent context data from the caller. Context is
keyed by tool name and persists across turns.
Declaring contextSchema
const weatherTool = tool({
name: 'get_weather',
description: 'Get weather for a location',
inputSchema: z.object({
location: z.string(),
}),
outputSchema: z.object({
temperature: z.number(),
}),
// Declare what context this tool needs
contextSchema: z.object({
apiKey: z.string(),
units: z.enum(['celsius', 'fahrenheit']),
}),
execute: async (params, context) => {
// Access this tool's own context via local
const { apiKey, units } = context.local;
const weather = await fetchWeather(
params.location,
apiKey,
units,
);
return { temperature: weather.temp };
},
});
Providing Context in callModel
Pass context keyed by tool name:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather in Tokyo?',
tools: [weatherTool, dbTool] as const,
// Static context — keyed by tool name
context: {
get_weather: { apiKey: 'sk-...', units: 'celsius' },
db_query: { connectionString: 'postgres://...' },
},
});
Dynamic Context
Use an async function for one-time initialization
that needs to fetch data:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather?',
tools: [weatherTool] as const,
// Resolved once at turn 0 to seed the store
context: async () => ({
get_weather: {
apiKey: await fetchApiKey(),
units: 'celsius',
},
}),
});
resolveContext runs once at turn 0 to seed the
context store. For per-turn mutations, use
setContext() inside your tool’s execute function.
Mutating Context with setContext
Tools can update their own context using setContext().
Changes persist across turns via the shared store and
are visible immediately — context.local is a live
getter that always reads the latest values:
const authTool = tool({
name: 'auth',
inputSchema: z.object({ action: z.string() }),
contextSchema: z.object({
token: z.string(),
refreshCount: z.number(),
}),
execute: async (params, context) => {
const { token } = context.local;
if (isExpired(token)) {
const newToken = await refreshToken(token);
// Mutate own context — persists to next turn
context.setContext({
token: newToken,
refreshCount:
context.local.refreshCount + 1,
});
}
return { success: true };
},
});
Observing Context Changes
Use getContextUpdates() on ModelResult to observe
context mutations in real time:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Authenticate and fetch data',
tools: [authTool] as const,
context: {
auth: { token: 'initial', refreshCount: 0 },
},
});
for await (const snapshot of result.getContextUpdates()) {
console.log('Context changed:', snapshot);
// { auth: { token: 'new-token', refreshCount: 1 } }
}
Shared Context
Use sharedSchema on tool() and sharedContextSchema
on callModel to share typed state across tools:
const SharedContextSchema = z.object({
_sessionId: z.string().optional(),
});
const execTool = tool({
name: 'sandbox_exec',
inputSchema: z.object({ command: z.string() }),
sharedSchema: SharedContextSchema,
execute: async (input, ctx) => {
// Read shared state set by any tool
const sid = ctx.shared._sessionId;
const session = await connect(sid);
// Write shared state for other tools
ctx.setSharedContext({ _sessionId: session.id });
return await session.exec(input.command);
},
});
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Run a command',
tools: [execTool] as const,
sharedContextSchema: SharedContextSchema,
context: {
shared: { _sessionId: 'existing-session' },
sandbox_exec: {},
},
});
context.local is scoped to one tool.
context.shared is visible to all tools and persists
across turns. Pass the same sharedSchema to each tool
for typed access, and sharedContextSchema to
callModel for runtime validation.
callModel automatically executes tools and handles multi-turn conversations. When the model calls a tool, the SDK executes it, sends the result back, and continues until the model provides a final response.
Automatic Execution Flow
When you provide tools with execute functions:
import { OpenRouter, tool } from '@openrouter/agent';
import { z } from 'zod';
const weatherTool = tool({
name: 'get_weather',
inputSchema: z.object({ location: z.string() }),
outputSchema: z.object({ temperature: z.number() }),
execute: async ({ location }) => {
return { temperature: await fetchTemperature(location) };
},
});
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather in Paris?',
tools: [weatherTool],
});
// getText() waits for all tool execution to complete
const text = await result.getText();
// "The weather in Paris is 18°C."
Execution Sequence
- Model receives prompt and generates tool call
- SDK extracts tool call and validates arguments
- Tool’s execute function runs
- Result is formatted and sent back to model
- Model generates final response (or more tool calls)
- Process repeats until model is done
Controlling Execution Rounds
Limit the maximum number of tool execution rounds:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Research this topic thoroughly',
tools: [searchTool, analyzeTool],
maxToolRounds: 3, // Stop after 3 rounds of tool execution
});
Setting maxToolRounds: 0 disables automatic execution - you get raw tool calls.
Use a function for dynamic control:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Research and analyze',
tools: [searchTool],
maxToolRounds: (context) => {
// Continue if under 5 turns
return context.numberOfTurns < 5;
},
});
The function receives TurnContext and returns true to continue or false to stop.
Get all tool calls from the initial response (before auto-execution):
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is the weather in Tokyo and Paris?',
tools: [weatherTool],
maxToolRounds: 0, // Don't auto-execute
});
const toolCalls = await result.getToolCalls();
for (const call of toolCalls) {
console.log(`Tool: ${call.name}`);
console.log(`ID: ${call.id}`);
console.log(`Arguments:`, call.arguments);
}
Stream tool calls as they complete:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Check weather in multiple cities',
tools: [weatherTool],
maxToolRounds: 0,
});
for await (const toolCall of result.getToolCallsStream()) {
console.log(`Received tool call: ${toolCall.name}`);
// Process each tool call as it arrives
const weatherResult = await processWeatherRequest(toolCall.arguments);
console.log('Result:', weatherResult);
}
Stream both argument deltas and preliminary results:
const searchTool = tool({
name: 'search',
inputSchema: z.object({ query: z.string() }),
eventSchema: z.object({ progress: z.number(), status: z.string() }),
outputSchema: z.object({ results: z.array(z.string()) }),
execute: async function* ({ query }) {
yield { progress: 25, status: 'Searching...' };
yield { progress: 50, status: 'Processing...' };
yield { progress: 75, status: 'Ranking...' };
yield { progress: 100, status: 'Complete' };
return { results: ['result1', 'result2'] };
},
});
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Search for TypeScript tutorials',
tools: [searchTool],
});
for await (const event of result.getToolStream()) {
switch (event.type) {
case 'delta':
// Raw argument delta from the model
process.stdout.write(event.content);
break;
case 'preliminary_result':
// Progress from generator tool
console.log(`Progress: ${event.result.progress}% - ${event.result.status}`);
break;
}
}
Event Types
| Type | Description |
|---|
delta | Raw tool call argument chunks from model |
preliminary_result | Progress events from generator tools (intermediate yields) |
When using getFullResponsesStream(), you can also receive tool.result events that fire when a tool execution completes:
for await (const event of result.getFullResponsesStream()) {
switch (event.type) {
case 'tool.preliminary_result':
// Intermediate progress from generator tools
console.log(`Progress (${event.toolCallId}):`, event.result);
break;
case 'tool.result':
// Final result when tool execution completes
console.log(`Tool ${event.toolCallId} completed`);
console.log('Result:', event.result);
// Access any preliminary results that were emitted during execution
if (event.preliminaryResults) {
console.log('All progress events:', event.preliminaryResults);
}
break;
}
}
type ToolResultEvent<TResult = unknown, TPreliminaryResults = unknown> = {
type: 'tool.result';
toolCallId: string;
result: TResult;
timestamp: number;
preliminaryResults?: TPreliminaryResults[];
};
The tool.result event provides the final output from tool execution along with all intermediate preliminaryResults that were yielded during execution (for generator tools). This is useful when you need both real-time progress updates and a summary of all progress at completion.
When the model calls multiple tools, they execute in parallel:
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Get weather in Paris, Tokyo, and New York simultaneously',
tools: [weatherTool],
});
// All three weather calls execute in parallel
const text = await result.getText();
For tools without execute functions:
const confirmTool = tool({
name: 'send_email',
description: 'Send an email (requires confirmation)',
inputSchema: z.object({
to: z.string().email(),
subject: z.string(),
body: z.string(),
}),
execute: false, // Manual handling
});
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Send an email to alice@example.com',
tools: [confirmTool],
maxToolRounds: 0,
});
const toolCalls = await result.getToolCalls();
for (const call of toolCalls) {
if (call.name === 'send_email') {
// Show confirmation UI
const confirmed = await showConfirmDialog(call.arguments);
if (confirmed) {
await sendEmail(call.arguments);
}
}
}
Execution Results
Access execution metadata through getResponse():
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'What is 2+2 and the weather in Paris?',
tools: [calculatorTool, weatherTool],
});
const response = await result.getResponse();
// Response includes all execution rounds
console.log('Final output:', response.output);
console.log('Usage:', response.usage);
Error Handling
Errors in execute functions are caught and sent back to the model:
const riskyTool = tool({
name: 'risky_operation',
inputSchema: z.object({ input: z.string() }),
outputSchema: z.object({ result: z.string() }),
execute: async (params) => {
if (params.input === 'fail') {
throw new Error('Operation failed: invalid input');
}
return { result: 'success' };
},
});
const result = openrouter.callModel({
model: 'openai/gpt-5-nano',
input: 'Try the risky operation with "fail"',
tools: [riskyTool],
});
// Model receives error message and can respond appropriately
const text = await result.getText();
// "I tried the operation but it failed with: Operation failed: invalid input"
Validation Errors
Invalid tool arguments are caught before execution:
const strictTool = tool({
name: 'strict',
inputSchema: z.object({
email: z.string().email(),
age: z.number().min(0).max(150),
}),
execute: async (params) => {
// Only runs with valid input
return { valid: true };
},
});
Graceful Error Handling
Handle errors gracefully in execute functions:
const robustTool = tool({
name: 'fetch_data',
inputSchema: z.object({ url: z.string().url() }),
outputSchema: z.object({
data: z.unknown().optional(),
error: z.string().optional(),
}),
execute: async (params) => {
try {
const response = await fetch(params.url);
if (!response.ok) {
return { error: `HTTP ${response.status}: ${response.statusText}` };
}
return { data: await response.json() };
} catch (error) {
return { error: `Failed to fetch: ${error.message}` };
}
},
});
Best Practices
Descriptive Names and Descriptions
// Good: Clear name and description
const tool1 = tool({
name: 'search_knowledge_base',
description: 'Search the company knowledge base for documents, FAQs, and policies. Returns relevant articles with snippets.',
// ...
});
// Avoid: Vague or generic
const tool2 = tool({
name: 'search',
description: 'Searches stuff',
// ...
});
Schema Descriptions
Add .describe() to help the model understand parameters:
const inputSchema = z.object({
query: z.string().describe('Natural language search query'),
maxResults: z.number()
.min(1)
.max(100)
.default(10)
.describe('Maximum number of results to return (1-100)'),
dateRange: z.enum(['day', 'week', 'month', 'year', 'all'])
.default('all')
.describe('Filter results by time period'),
});
Design tools to be safely re-executable:
const createUserTool = tool({
name: 'create_user',
inputSchema: z.object({
email: z.string().email(),
name: z.string(),
}),
execute: async (params) => {
// Check if user exists first
const existing = await findUserByEmail(params.email);
if (existing) {
return { userId: existing.id, created: false };
}
const user = await createUser(params);
return { userId: user.id, created: true };
},
});
Timeout Handling
Wrap long-running operations:
const longRunningTool = tool({
name: 'process_data',
inputSchema: z.object({ dataId: z.string() }),
execute: async (params) => {
const timeoutMs = 30000;
const result = await Promise.race([
processData(params.dataId),
new Promise((_, reject) =>
setTimeout(() => reject(new Error('Operation timed out')), timeoutMs)
),
]);
return result;
},
});
Next Steps