Overview

Nora provides end-to-end visibility into modern LLM applications by automatically tracking LLM calls, agent workflows, and tool executions.

Modern AI systems are no longer simple prompt–response interactions. They involve multi-step workflows, agent coordination, routing decisions, tool usage, and streaming responses.

Nora makes these executions inspectable and debuggable by capturing not only what happened, but why it happened — without invasive or manual instrumentation.

Key Features

• Automatic tracking of OpenAI, Anthropic, and Google Gemini API calls

• First-class decision tracking (tool selection, routing, RAG choices)

• Full async/await and streaming support

• Trace groups for organizing multi-step logical workflows

• Custom function tracking via lightweight decorators

• Session- and user-level trace correlation

• Automatic tool and function calling detection

• Token usage, latency, and error capture

• Multiple span types for precise operation classification

Built for Production AI Systems

Nora is designed for real-world production environments:

• Low-overhead batching and async flushing

• Safe for FastAPI, background tasks, and streaming servers

• Works with synchronous, async, and parallel execution models

• Captures partial failures and errors as structured traces

This makes Nora suitable not only for experimentation, but for monitoring real user traffic at scale.

Supported Providers

• OpenAI: GPT-5, GPT-4, GPT-3.5 (synchronous, asynchronous, streaming)

• Anthropic: Claude Opus, Sonnet (synchronous, asynchronous, streaming)

• Google Gemini: Gemini Pro, Flash (synchronous, asynchronous, streaming)