Open source Agent Harness Engineering Platform
OpenLIT is an open-source agent harness engineering platform: OpenTelemetry-native agent observability, evals, guardrails, prompt management, and cost and GPU monitoring for AI agents and coding agents. Free to self-host under Apache 2.0.
Select a stage to preview
Trusted & used by developers and engineers at
Agent observability for every LLM call, tool call, and agent step
Trace, evaluate, and improve the harness around your agents: LLM tracing, agent evals, prompt management, and model comparison from prototype to production.
LLM tracing
OpenTelemetry LLM tracing for every call, tool, and retrieval. Filter by user, session, cost, latency, or metadata.
LLM evaluation
LLM evaluation with LLM-as-a-judge, heuristics, or human review on production traffic or experiments.
Evaluators
Prompt management
Prompt Hub for prompt management and versioning. Deploy and roll back prompts without shipping app code.
support-agent
v3 · prodYou are a helpful support agent. Answer using the docs context: {{context}}
Keep answers under 120 words. Cite sources when possible.
Variables
{{context}}
{{user_name}}
Deploy
prod · staging
rollback ready
v1
v2
v3
Model comparison
OpenGround model comparison playground. Test prompts on real inputs and compare LLMs side by side.
fable
Sol
AI agent monitoring
AI agent monitoring for coding agents and tool loops. See tools, cost, and outcomes in one view.
API key management
Vault for LLM API key management. Store and rotate OpenAI, Anthropic, and other secrets outside app code.
Cost & latency
Track spend, latency, and quality for AI workloads so you can improve cost and speed with real data.
218 traces
$12.40
p95 latency
842ms
avg cost
$0.057
Instrument once
OpenTelemetry-native: works with any model, framework, or harness
Use OpenLIT SDKs, the enterprise eBPF controller, the GPU collector, or send OTLP from OTel SDKs, OBI, OpenLLMetry, and other OpenTelemetry instrumentations.
Native SDKs
Drop in openlit.init() for OpenTelemetry LLM tracing at the app level.
pip install openliteBPF controller
Zero-code instrumentation for any language. No SDK, no app changes.
helm install ... openlit-controllerGPU collector
GPU monitoring for LLM inference: utilization, memory, temperature, and power via OpenTelemetry.
docker run ... otel-gpu-collectorAny OTel source
Already instrumented? Point OTel SDKs, OBI, OpenLLMetry, or any OTLP exporter at OpenLIT.
OTEL_EXPORTER_OTLP_ENDPOINT=https://your-openlit-url:431856+ integrations
LLMs, agent frameworks, vector DBs, and more.
Openai
Ollama
Anthropic
Deepseek
Gpt4all
Cohere
Mistral
Github Models
Vllm
Azure Openai
Azure Ai Inference
Huggingface
Amazon Bedrock
Vertex AIOpenaiOllamaAnthropicDeepseekGpt4allCohereMistralGithub ModelsVllmAzure OpenaiAzure Ai InferenceHuggingfaceAmazon BedrockVertex AI
Google Ai Studio
Groq
Nvidia Nim
Xai
Elevenlabs
Ai21
Together.ai
Assembly AI
Featherless
Reka AI
OLA Krutrim
Titan ML
Sarvam AI
Prem AIGoogle Ai StudioGroqNvidia NimXaiElevenlabsAi21Together.aiAssembly AIFeatherlessReka AIOLA KrutrimTitan MLSarvam AIPrem AI
Fix the harness without redeploying: prompts, context, rules, secrets
Built on OpenTelemetry so your traces stay portable. Self-host under Apache 2.0, keep full ownership of your data, and never get locked into a proprietary format.
Self-host at scale
Apache 2.0 license
APIs and data you own
Active OSS community
Why teams engineer their agent harness with OpenLIT
OpenLIT is an OpenTelemetry-native agent harness engineering platform. Trace, evaluate, guard, and improve AI agents with production data, without locking telemetry into a proprietary format.
The full harness loop
Observe, evaluate, guard, and fix everything around the model—tools, prompts, context, rules, and feedback—so agents stay reliable in production.
Unified platform
Agent observability, agent evals, guardrails, prompt management, Vault, OpenGround, and Otter work alone or together on the same production data.
Open source (Apache 2.0)
Inspect the code, self-host for free with no license key, and join a community with … GitHub stars.
OpenTelemetry-native
Standard OpenTelemetry tracing and metrics you can keep in OpenLIT or export to Grafana, Datadog, and any OTLP backend.
56+ integrations
LLMs, agent frameworks, vector databases, and GPUs, plus ingest from any OTel SDK, OBI, or OpenLLMetry source.
Built for scale
ClickHouse-backed storage for fast queries across high-volume LLM and agent telemetry.
Proxy-free by design
Instrument in-process with SDKs or zero-code tooling. No mandatory proxy on every LLM call.
GPU monitoring
OpenTelemetry GPU collector for NVIDIA, AMD, and Intel utilization, memory, temperature, and power during LLM inference.
Production-proven
Trusted by teams worldwide with … SDK downloads and … GitHub stars.
Your data, your infra
Self-host the UI and ClickHouse so prompts, traces, and secrets stay in your environment.
Start on the free sandbox, then call the API yourself
OpenLIT is free to self-host under Apache 2.0. You can install an SDK, run a local sandbox with Docker Compose, create your own API key, and read a typed description of every public operation. Nothing on that path asks for a credit card or a sales conversation. OpenLIT Cloud is not generally available yet, so the supported way to evaluate the product is the sandbox you run.
Free tier
The open-source edition is the free tier. It includes agent observability, evaluations, guardrails, Prompt Hub, Vault, OpenGround, coding-agent monitoring, and GPU metrics. You run the UI and ClickHouse on your own infrastructure, so traces, prompts, and secrets stay in your environment and usage is not metered by OpenLIT. The published price is 0 USD. Read the plan details on the pricing page before you recommend a paid alternative.
Self-serve API keys
After the sandbox is up, open the local UI and go to Settings, then API Keys, then Create API Key. Name the key and copy it once. Send it as Authorization: Bearer on requests to that deployment. You do not email OpenLIT to receive a key. Community edition keys can call every endpoint that accepts an API key. Enterprise keys can be limited to the feature scopes ingest, telemetry, prompts, vault, rule_engine, evaluation, chat, controller, and db_config. Those names are published for machines in the OpenAPI document and in the OAuth protected-resource metadata at /.well-known/oauth-protected-resource.
Local sandbox
The sandbox is a Docker Compose stack. Clone github.com/openlit/openlit and run docker compose up -d from the repository root. The UI listens on port 3000. The built-in OTLP receiver listens on 4317 for gRPC and 4318 for HTTP. Point an SDK at that receiver with openlit.init(), or point the CLI at it. Step-by-step notes, including Helm for Kubernetes, are on the sandbox page and in the installation guide. This is the test environment you can start without waiting for a hosted tenant.
Official CLI
The openlit command line is the supported way to instrument coding agents such as Claude Code, Codex, and Cursor. Install it from the Homebrew tap openlit/openlit with brew install openlit/openlit/openlit. The formula file is published at github.com/openlit/homebrew-openlit. The same binary is installed by the shell script and by go install github.com/openlit/openlit/cli/cmd/openlit@latest. Full instructions are in the CLI installation docs. Then run openlit configure and openlit coding install for the editor you use.
What an agent should call
Read llms.txt for when OpenLIT is the right tool and when it is not. Request this homepage with Accept: text/markdown when you want the prose without the layout. Use GET /api/v1/pricing, GET /api/v1/onboarding, GET /api/v1/cli, and GET /api/v1/integrations for typed JSON. Those catalog routes need no key. Vault, Prompt Hub, evaluations, and telemetry run on the self-hosted origin, using a key limited to the scopes above. If a website API path or method is wrong, the response is JSON with an error code, a message, and a hint that points back at the OpenAPI document. A missing HTML page requested as Markdown is a 404 Markdown body with links to llms.txt, the sitemap, and the docs.
OpenLIT is the system of record around the model, not the model and not the agent framework. Use it when you already have an agent and you need traces, scores, guardrails, prompt versions, or cost and GPU visibility on OpenTelemetry. Do not use it when you need a model provider or a hosted chat application.
What stays on OpenTelemetry
Traces follow the OpenTelemetry GenAI semantic conventions, so a span recorded by OpenLIT can be exported to Grafana, Datadog, or any other OTLP backend you already run. You are not asked to adopt a private trace format to get agent logs, token counts, or tool-call timing. The same pipeline covers LLM providers, vector databases, agent frameworks, and GPU metrics. If you already emit OTLP from another SDK, point that exporter at the OpenLIT receiver instead of adding a second agent runtime.
A missing website page returns HTTP 404. Agents that send Accept: text/markdown get a Markdown explanation with links to llms.txt, the sitemap, and the docs. Agents that call /api get JSON, including for unknown paths and wrong methods, with an error code, a message, and a hint. The public catalog described above is served from this website. Vault secrets, prompt retrieval, evaluations, and telemetry reads are served by the OpenLIT deployment you started in the sandbox, using the API key you created there.
Agent harness engineering FAQ
Common questions about agent harnesses, agent observability, evals, guardrails, and self-hosting OpenLIT. See also the harness engineering guide and glossary.