Start
NeuronLens documentation
How NeuronLens checks AI answers, agents and models from the inside, and how to connect it.
Start in five minutes
Products
Guides
Integrations and hooksHow the Claude Code hook, MCP gateway, LiteLLM guardrail, OpenTelemetry, the SDK, LangGraph and Google ADK each connect, what they can stop, and how to set them up.Read more →Screen a model before you deploy itA short Python script that scans a fine-tune or merge and stops if its safety was stripped.Read more →
Reference
NeuronLens checks AI from the model's internals: the activations an answer produces and the weights a model is made of. A text-only check sees what a model says. NeuronLens reads what the model represents while it says it.
| Product | You send | You get back |
|---|---|---|
| Agent Lens | Your app's answers, or each step your agent is about to take | A risk per failure mode (deception, goal drift, data leak, reasoning error, failed check, omission, …) and a verdict: allow, escalate or deny |
| Integrity Lens | A Hugging Face model, public or private | Whether the model's safety training has been stripped (no issue, review, tampered), whether it is still the version you approved, and a signed report |
Make your first call
- Sign in at neuronlens.com and create a key under Profile → API keys.
- Score an answer:
curl https://api.neuronlens.com/v1/score \
-H "Authorization: Bearer $NEURONLENS_API_KEY" -H "Content-Type: application/json" \
-d '{"prompt": "What is our refund window?",
"response": "Refunds are available within 14 days of purchase with a receipt.",
"agent": "support-bot"}'
- Read
verdictandprobes, then open Monitor in the console to see the answer. The Agent Lens quickstart takes five minutes.
Connect your agent
| Where your agent runs | Start here |
|---|---|
| Your own code | Check agent steps: one HTTP call before each tool |
| Claude Code | Claude Code hook |
| MCP tools | MCP gateway |
| LiteLLM proxy | LiteLLM guardrail |
| LangGraph, LangChain, Google ADK | Framework middleware |
| Already traced with OpenTelemetry | OpenTelemetry |
How it works
- Agent Lens reads your conversation through an open-weight model (gpt-oss-20b) and runs a probe per failure mode on its internal activations. A model that writes a misleading answer, or an agent about to act against a failed check, represents it differently inside, even when the text reads well.
- Integrity Lens compares a model's weights and late-layer refusal signal with the clean base it was built from. Abliteration and similar edits leave a measurable gap. Your organization approves a version of each model, and every later scan is checked against it.
Where it runs
- API:
https://api.neuronlens.com, with your key asAuthorization: BearerorX-API-Key. See Authentication. - Console: app.neuronlens.com: Agent Lens's Connect, Monitor, Eval and Settings (Pro), and Integrity Lens scans, approvals and reports for your organization.
Scans in the web app count toward the same daily allowance as the API. See Limits and usage.