Self-hosted LLM observability with built-in air-gap mode. Quality scoring, agent waterfalls, prompt debugger, GDPR-native compliance workflows. Free for production self-hosting (BSL 1.1).
🔗 Live demo (no signup) · 📦 pip install agentlens-monitor · 🐳 Self-host in 5 min
Most LLM observability tools force you to choose: ship your prompts to a US-hosted SaaS, or build it yourself. AgentLens gives you a third option — a production-grade self-hosted platform that runs entirely in your infrastructure, with one Docker command.
You get:
- 🎯 Quality scoring + hallucination flags — heuristic by default, optional LLM-as-judge
- 🪵 Agent waterfall debugging — multi-step traces with span-level timing and cost
- 🔍 Prompt-level inspection — every LLM call, with full input/output and metadata
- 💰 Cost + budget observability — per-model, per-feature, with daily budget alerts
- 🛡️ GDPR-native compliance — export, retention policy, audit log, right-to-erasure
- 🔐 Air-gap mode — single env var disables every external integration. tcpdump-verifiable.
| AgentLens | LangSmith | Langfuse | Helicone | |
|---|---|---|---|---|
| Self-hosted | ✅ default | ❌ cloud-only | ✅ | ✅ |
| Air-gap mode (no outbound) | ✅ baked-in | ❌ | ||
| GDPR-native + EU-hosted cloud | ✅ | ❌ US | ❌ US | |
| Heuristic scoring (no judge calls) | ✅ default | ❌ paid LLM judge | ❌ paid LLM judge | ❌ |
| Agent waterfall traces | ✅ | ✅ | ✅ | ❌ |
| BSL/MIT open-core | ✅ | ❌ proprietary | MIT | MIT |
| Production cost (mid-team) | €299–2,999/mo | $39–500/seat | $$ | $$ |
Picking guide: Pick AgentLens if you're in regulated industry, EU-based, or air-gap matters to you. Pick LangSmith if you want a fully managed cloud and don't mind US data residency. Pick Langfuse if you want a free MIT-only path. Pick Helicone for simple proxy-style logging.
One-line installer (recommended):
curl -fsSL https://www.agentlens.one/install | bashOR build from source (security-conscious teams):
git clone https://github.com/Soufianeazz/agentlens && cd agentlens
cp .env.airgap.example .env
echo "ADMIN_TOKEN=$(openssl rand -hex 32)" >> .env
docker compose up -d --build→ Open http://localhost:8000 for the dashboard.
Add the SDK to your code:
pip install agentlens-monitorimport agentlens
agentlens.init(api_url="http://localhost:8000/ingest", api_key="al_...")
agentlens.patch_openai() # auto-track every OpenAI call
agentlens.patch_anthropic() # auto-track every Anthropic callfrom agentlens import trace_agent
with trace_agent("research_agent", input="Research market trends") as trace:
with trace.span("retrieve_context", span_type="retrieval") as s:
s.set_output("found 12 documents")
with trace.span("reason_and_draft", span_type="llm", model="gpt-4o") as s:
s.set_output("Draft ready")
s.set_tokens(1200)
s.set_cost(0.009)
trace.set_output("Report complete")→ Every span shows up in the agent waterfall view with timing, cost, and full I/O.
I'm Soufian, a solo founder. I kept seeing teams stuck between two bad choices for LLM observability: ship sensitive prompts to a US SaaS, or build it themselves and abandon it after 6 weeks. The third option — a production-grade self-hostable platform with EU compliance baked in — didn't exist.
So I built it. AgentLens is what I'd want to run on my own infra: one Docker command, no telemetry, GDPR-clean, and a clear paid tier for teams that want managed operation.
If this resonates: ⭐ a star helps the project show up in search, and tells me to keep building.
| Plan | Price | What's included |
|---|---|---|
| Free | €0 | Ingest API + Basic Stats (24h) + Prompt Debugger preview |
| Starter | €299/mo | + Prompt Debugger |
| Team | €999/mo | + Agent Debugger + Advanced Analytics |
| Scale | €2,999/mo | + Compliance / GDPR workflows + Audit log |
| Enterprise | €5,000+/mo | + Private Deploy + SLA + Security Package |
Self-hosting under BSL is free for any internal production use — paid plans add server-side feature tiers (Prompt/Agent Debugger, Compliance workflows, Audit Log) plus managed cloud, support, and security artifacts.
Qualified teams can request a 14-day full-feature pilot before committing.
- Python SDK (
agentlens/, on PyPI) → MIT. Embed in any product. No restrictions. - Server (everything else in this repo) → Business Source License 1.1. Free for self-hosted production use. Converts to Apache 2.0 on 2030-05-11.
The BSL only restricts offering AgentLens itself as a competing hosted SaaS to third parties. Internal production use, modification, and redistribution within your org are explicitly permitted. See LICENSING.md for the plain-language guide.
| Endpoint | Purpose |
|---|---|
POST /ingest |
Log an LLM call |
GET /requests/stats |
24h KPIs (calls, quality, cost) |
GET /requests/trend |
Hourly quality trend |
GET /requests/worst |
Worst-quality responses |
GET /requests/cost-quality |
Cost vs quality per model |
GET /debug/requests |
Full call inspector |
POST /traces + lifecycle |
Agent waterfall traces |
GET /compliance/export |
CSV/JSON export |
POST /compliance/retention |
Auto-delete policy |
GET /compliance/audit-log |
Audit events (admin) |
Full docs in api/main.py and docs/SELF_HOST.md.
Use agentlens.one for managed rollout, buyer-ready trust artifacts, and plan-based feature access. Hosted in Frankfurt (EU), GDPR-native by design.
docs/SELF_HOST.md— full self-host guideLICENSING.md— license model in plain English- Source of truth + deploy path:
docs/SOURCE_OF_TRUTH.md - Pre-call deploy checklist:
docs/DEPLOY_CHECKLIST_PRECALL.md - Smoke tests after deploy:
docs/SMOKE_TEST_RUNBOOK.md
Recommended env hardening for production:
CORS_ALLOWED_ORIGINS=https://www.agentlens.one,https://agentlens.one
AGENTLENS_AIRGAP=1 # for fully isolated deployments
- 🐛 Bug reports + feature requests: GitHub Issues
- 💬 Questions / pilot requests: soufian.azzaoui48@gmail.com
- ⭐ Star the repo if you'd like to see this category mature
If you're shipping AgentLens in production, I'd love to hear about it — and a quote / case study (with logo) gets you a discount on the Scale tier.