Tool calling

Define tools#

import { generate, tool } from 'edgewise';
import { z } from 'zod';

const tools = {
  getWeather: tool({
    description: 'Current weather for a city',
    input: z.object({ city: z.string() }),
    execute: async ({ city }) => fetchWeather(city),
  }),
  addTodo: tool({
    description: 'Add an item to the to-do list',
    input: z.object({ title: z.string(), due: z.string().optional() }),
    execute: async (todo) => db.todos.add(todo),
  }),
};

const { text, toolCalls, toolResults } = await generate({ model: 'text:default', input: 'Remind me to call Ana tomorrow', tools });

The loop#

When the model calls tools, Edgewise validates each call's input with its schema, runs execute, adds the results to the conversation, and calls the model again, up to maxSteps (default 3). An invalid call is sent back to the model as an error instead of being run. The run ends when the model answers in text, or with finishReason: 'tool-calls' when steps run out.

Tools without execute are returned in toolCalls for you to handle.

Approve before running#

await generate({
  model: 'text:default',
  input,
  tools,
  approve: async (call) => call.name !== 'deleteAll' && confirm(`Run ${call.name}?`),
});

A declined call is reported to the model as declined.

Events#

const run = generate({ model: 'text:default', input, tools });
for await (const e of run.events) {
  if (e.type === 'tool-call') log(`→ ${e.call.name}`, e.call.input);
  if (e.type === 'tool-result') log(`← ${e.result.name}`, e.result.output);
}

Models and formats#

Each model family writes tool calls its own way. Edgewise speaks each format:

Model	Format
LFM2 and LFM2.5	`<
Qwen3	Hermes: `<tool_call>{"name": …}</tool_call>`
FunctionGemma	`<start_function_call>call:fn{…}<end_function_call>`

Small models call tools best when the tool names and descriptions are plain and specific, and when there are fewer than ten tools. functiongemma-270m and lfm2-1.2b-tool are trained for app actions.