Open Source

Trace every LLM call.
One line of code.

Drop in @farzanhossans/agentlens. OpenAI, Anthropic, Gemini, Cohere, Mistral — every call traced automatically. Costs, errors, session replay included. No client wrappers, no code changes inside your call sites.

MIT licensed  ·  Self-host in 2 minutes  ·  No vendor lock-in

terminal

AI agents fail silently.
You deserve to know why.

🔴

Your agent returned the wrong answer. You have no idea which LLM call went wrong.

💸

Your OpenAI bill hit $2,000. You can't explain which agent or feature caused it.

🐛

A bug in production. You can't reproduce it because you have no trace of what the model saw.

🤯

5 agents, 3 models, streaming everywhere. Something broke. Good luck finding which call.

Two ways to integrate.
Start with one line.

The universal SDK covers every major provider in one line. Drop down to manual tracing when you need custom span hierarchies.

Option 2

Manual — full control

Wrap any code in AgentLens.trace() for custom spans, metadata, and parent/child hierarchies.

// Full control over spans

AgentLens.trace('classify', (span) => {
  span.setMetadata('user', id)
  // your LLM calls here
})

Everything you need to
debug AI agents.

Universal SDK

One line traces every provider — OpenAI, Anthropic, Gemini, Cohere, Mistral. Network-layer interception works with axios, got, node-fetch, or any client.

🔍

Trace Viewer

Full input/output timeline with parent/child span hierarchy. See exactly what the model received and returned.

💰

Cost Analytics

Token usage and dollar cost by model, agent, and date. Instant aggregations across millions of spans. Monthly budget tracking.

🚨

Smart Alerts

Error rate, cost spike, P95 latency, failure count. Real-time metric evaluation. Notify via Slack, email, or webhook.

Live Feed

Watch agent calls in real-time via WebSocket. See traces as they happen in production.

Session Replay

Step through any past agent run. Group traces by session to see multi-turn conversations.

🔒

PII Scrubbing

Emails, API keys, SSNs, credit cards auto-masked before data leaves your infrastructure. GDPR ready.

🌊

Streaming-safe

SSE streams are tapped via ReadableStream.tee() — the bytes you read are byte-identical to the original. Zero latency added.

🔗

Trace Grouping

Group multiple LLM calls into one trace with optional headers. See parent/child span hierarchies for multi-step agent runs.

🧩

Error Clustering

Similar failures auto-grouped with count badges and affected models. Spot patterns instantly instead of scrolling through logs.

🏠

Self-Hostable

One docker compose command. Your data stays on your servers. No third-party accounts required.

Your data. Your servers.
2 minutes to deploy.

  • Full stack in one docker compose command
  • No Vercel, no Cloudflare, no third-party accounts
  • PostgreSQL, Redis, Elasticsearch — search, analytics, and auto-cleanup built in
  • MIT licensed — free forever, no usage limits
  • Custom domains and SSL with any reverse proxy
View on GitHub
terminal
# Clone and configure
git clone https://github.com/farzanhossan/agentlens
cd agentlens/infra
cp .env.prod.example .env

# Generate secrets
openssl rand -hex 32  # paste as JWT_SECRET
openssl rand -hex 32  # paste as HMAC_SECRET

# Start everything
docker compose -f docker-compose.prod.yml up -d --build

# Dashboard  http://localhost:4021
# API        http://localhost:4020

Why developers choose AgentLens.

Feature AgentLens LangSmith Langfuse Helicone
Universal SDK — one line, every provider ❌ LangChain only ⚠️ Per-provider ❌ Proxy only
Any language / framework ❌ LangChain only ⚠️
PII scrubbing built-in
Session replay
Self-host (open source) ✅ MIT
Trace grouping (multi-call)
Real-time live feed
Free tier Unlimited (self-host) 5k traces 50k events 10k requests

Free forever. Self-hosted.

Start debugging your AI agents today.

Deploy in 2 minutes. Free forever. No vendor lock-in.

MIT licensed  ·  Docker only  ·  No third-party accounts needed