Trace LLM calls

Record every model call with tokens, latency, errors and cost.

Tracing records model calls. To govern what an agent does, also govern its tool calls.

Wrap your provider's client

The second argument names the agent. It groups traces in the console and is the name used for the kill switch and policies.

import OpenAI from 'openai';
import { Leap360 } from 'leap360';

const leap = new Leap360({ apiKey: process.env.LEAP360_API_KEY });
const openai = leap.wrapOpenAI(new OpenAI(), 'search-agent');

Use the client that the wrap call returns. Calls made on the original client are not traced.

Zero config

import { autoInit } from 'leap360';

await autoInit(); // reads LEAP360_API_KEY

Any other provider

Wrap the call with traceCall. It times the call, captures errors and sends the trace.

const result = await leap.traceCall({
  model: 'mistral-large',
  provider: 'mistral',
  serviceName: 'mistral-agent',
  fn: async () => mistralClient.chat({ model: 'mistral-large', messages }),
});

Python

from leap360 import Leap360

leap = Leap360()  # reads LEAP360_API_KEY

leap.trace(
    model="amazon.nova-micro-v1:0",
    provider="bedrock",
    service_name="my-agent",
    latency_ms=312,
    usage={"prompt_tokens": 80, "completion_tokens": 50, "total_tokens": 130},
)

See the Python SDK reference for provider wrappers.

Kill switch

Stop one agent or the whole project, right away.

Console tour

What each screen of the console is for, with a picture of it.

On this page

Wrap your provider's client Zero config Any other provider Python