OpenLIT

Visit Website

Introduction

OpenLIT describes itself as an open-source agent harness engineering platform. Its view is that an AI agent is a model plus a harness, meaning the tools, context, prompts, memory, hooks, guardrails and feedback loops around the model, and that reliability comes from measuring and improving that harness. The loop it supports is run, observe, evaluate, fix the harness and verify.

Observability is built on OpenTelemetry rather than a proprietary format. A single openlit.init() call auto-instruments supported LLM providers, frameworks, vector databases and other AI infrastructure, exporting traces and metrics that follow the OpenTelemetry GenAI semantic conventions (gen_ai.*). You can filter traces by user, session, cost, latency or metadata, and connectors plug the data into ClickHouse, Tempo, Loki, Prometheus and more. The README lists more than 70 integrations, including the OpenAI Agents SDK, LangGraph and CrewAI.

Coding agents get their own treatment. A CLI installer gives an OpenTelemetry view of sessions from Claude Code, Cursor, Codex, Windsurf and OpenCode, explored in a Coding Agents dashboard, which is useful when you want to understand what an agent actually did.

On the testing side, OpenLIT runs LLM-as-a-judge and programmatic evaluations on production traffic or experiments and supports CI gates. SDK guardrails detect and block prompt injection and sensitive topics at runtime, and every guardrail decision is traced. OpenGround runs the same prompt across providers to compare cost, latency and quality before you ship.

It also includes prompt management kept separate from code, a Context store for reusable RAG content, a Vault for API keys, and an OpenTelemetry GPU collector for NVIDIA, AMD and Intel GPUs that reports utilisation, memory, power and temperature correlated with traces.

The source code is at openlit/openlit on GitHub under the Apache-2.0 licence, with about 3,000 stars; the most recent commit was on 10 November 2026.

OpenLIT fits teams that want LLM observability without adopting a new proprietary protocol. Because it speaks OpenTelemetry, traces from LLM calls can sit next to traces from the rest of the stack, which makes it easier to tell whether a slow response came from the model, a vector database or your own code.

A sensible starting point is to instrument one service, check that token usage and latency numbers match your provider’s bills and dashboards, and only then roll it out widely. Self-hosting keeps prompts and responses inside your infrastructure, which often matters for compliance reviews.

Watch costs as well as quality. Token usage dashboards make it obvious when a prompt change doubles spend, which is just as much a regression as a drop in answer quality.

It is also a good fit for platform teams that support several AI products, since one OpenTelemetry pipeline can serve all of them with consistent attributes and dashboards.

Launch and background: OpenLIT was posted to Hacker News as a Show HN on 26 April 2024.

Features

1. OpenTelemetry-native tracing: GenAI semantic conventions, one-line init.
2. 70+ integrations: Providers, frameworks and vector databases.
3. Coding-agent observability: Claude Code, Cursor, Codex, Windsurf and OpenCode sessions.
4. Evals: LLM-as-a-judge, heuristics and human review.
5. CI gates: Fail builds on eval thresholds.
6. Guardrails: Prompt injection and sensitive-topic detection, traced.
7. Prompt management: Versioned prompts outside your code.
8. OpenGround: Compare providers on cost, latency and quality.
9. Vault: Store LLM keys and fetch them at runtime.
10. GPU monitoring: NVIDIA, AMD and Intel GPU metrics.

Pros

1. No lock-in: Standard OpenTelemetry data you can send anywhere.
2. All-in-one: Observability, evals and guardrails in one self-hosted stack.
3. Free to use: Open source under the Apache-2.0 licence, so the core tool costs nothing to use.
4. Actively maintained: The repository had commits as recent as 10 November 2026.
5. Community: About 3,000 GitHub stars.

Cons

1. Self-hosting work: You run the stack and its database.
2. Breadth over depth: Specialist eval tools may offer more metrics.
3. Fast release cadence: Frequent 2.x releases.

Usage and Links (Example)

Run the platform and instrument an app:

git clone https://github.com/openlit/openlit.git
cd openlit && docker compose up -d     # UI at http://127.0.0.1:3000
pip install openlit

import openlit
openlit.init()   # auto-instruments supported LLM SDKs and frameworks

Resources:

• OpenLIT website
• GitHub repository

Pricing

OpenLIT is free and open source under the Apache-2.0 licence.

Recommended for?

1. Teams standardised on OpenTelemetry who want AI traces in the same pipeline.
2. Platform teams observing coding agents across developers.
3. Self-hosting shops that cannot send prompts to a SaaS.

Conclusion

OpenLIT is a strong open-source choice when you want LLM observability, evals and guardrails without leaving OpenTelemetry.

Begin with tracing, then add evals and guardrails on the riskiest paths.

On CheckOps, OpenLIT is filed under the Production monitoring & observability and Release & CI gates stages. Related tools listed here include Langfuse, Arize Phoenix, OpenLLMetry, Helicone, Opik and SigNoz. Before standardising on one, compare recent maintenance activity, the pricing model as your usage grows, and the integrations you rely on today. Information on this page was verified against the official website, documentation and repository on 11 November 2026.