NeuronLens
Browse the documentation
Agent Lens

Agent Lens vs Agent Lens Pro

The playground, the API and the Pro console. What each one does, who can use it, and which one you need.


Agent Lens comes in three parts. The playground is for trying it. The API is for putting it in your app or agent. The Pro console is for running it in production: watching the traffic, reviewing what was flagged, approving held steps and testing agents before release.

1 · TRYAgent Lens playgroundapp.neuronlens.com → Agent LensPaste a prompt. gpt-oss-20bwrites the answer, or paste yours10 probes: answers and agent runsNothing to set up10 runs a day2 · BUILDAgent Lens APIAny nl_ key from your profilePOST /v1/score for answersand for steps before they runIntegrations: Claude Code, MCP,LiteLLM, OTel, LangGraph, ADK1,000 runs a day3 · RUNAgent Lens Pro consoleapp.neuronlens.com/agent-lensMonitor every answer and runApprove or reject held stepsEval and signed audit reportsSettings, rules, workspacesPro account
The playground and the API are open to every account. The console is where the API's results show up, and it needs Pro.

Side by side

Agent Lens playgroundAgent Lens APIAgent Lens Pro console
Use it tosee what the probes say about one promptcheck answers and agent steps from your codewatch, review, approve and test in production
Whereapp.neuronlens.com → Agent Lenshttps://api.neuronlens.comapp.neuronlens.com/agent-lens
Who can use itanyone signed inanyone with an nl_ keyPro accounts
EndpointPOST /v1/agent/scorePOST /v1/score, plus the integration endpointsthe API's results, in the browser
Answersyes: deception, goal drift, data leak, reasoning error, plus cyber and hallucinationyes: deception, goal drift, data leak, reasoning errorevery scored answer, with the flagged sentence
Agent runsyes: a tool log scored as a whole, or a plan before it runs (omission, failed check, harmful I/O, injected intent)yes: each tool call before it runs, inside your agent or through an integrationevery run, step by step
Who writes the answergpt-oss-20b, or you paste your ownalways your model
Holds for a personnoyes (escalate, then poll the decision)Approve and Reject buttons
Storedno historyunder your workspace (or nothing, with store: false)searchable for your plan's retention
Review and labelsnoPOST /v1/labelsConfirmed failure / False alarm, precision per probe
Eval and audit reportsnonoyes
Thresholds, rules, alertsfixed defaultsuses your workspace's Settingsset them in Settings
Runs per day101,000uses the API's runs
Requests per minute10300

Which one do I need?

  • "I want to see whether this works on my kind of answer." Use the playground. Paste a real prompt and your model's answer, and read which probe fires and which words pushed it.
  • "I want to check every answer my bot gives." Use the API. Call /v1/score after your model answers, and act on the verdict.
  • "I want to stop an agent before it does something wrong." Use the API with an integration (Claude Code hook, MCP gateway, LiteLLM, LangGraph, ADK) or a direct /v1/score check before each tool.
  • "Someone has to review flags and approve held steps." You need the Pro console. Held steps wait in Monitor → Agents → Open issues until a person clicks Approve or Reject.
  • "We need to test an agent before release, or show an auditor." You need the Pro console's Eval: failure scenarios, results per step, and a signed report.

The playground

Open app.neuronlens.com, sign in, and pick Agent Lens in the sidebar. Type a prompt. Either let gpt-oss-20b write an answer, or paste your own model's answer to score it as given. The page shows the verdict, the risk for each probe, and the words that pushed each risk up.

The same thing works from code, with the same 10 runs a day:

curl https://api.neuronlens.com/v1/agent/score \
  -H "Authorization: Bearer $NEURONLENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Our fund lost 18% this year. A client asks: is my money safe with you?",
       "response": "Yes, your money is completely safe. There is no risk at all."}'

Leave out response and gpt-oss-20b writes the answer, which comes back as answer. The response has the verdict, each probe's risk, and loci: the tokens that pushed the score, each with its weight. Playground results are not stored in a workspace and do not appear in the console.

Probes in the playground. For a prompt: deception, goal drift, data leak, reasoning error, cyber (the request asks for attack help: malware, intrusion, exploits) and hallucination (the answer states something its source does not support; put the source in system). GET /v1/agent/probes lists every probe with "live" and its category.

Agent runs in the playground. Send the run's tool log as transcript ("tool: result" lines, up to 40) and the tools the agent was offered as offered. With response, the agent's report is scored as given; without it, the model writes the report from the log. With "before_action": true, the log ends in the steps the agent plans next ("tool: (planned next)"), and the plan is scored before anything runs:

{"prompt": "Handle the driver release of vehicle VH-7924, and make sure every required verification is completed before the final step.",
 "transcript": ["get_vehicle_record: (planned next)", "release_vehicle_to_driver: (planned next)"],
 "offered": ["verify_insurance_active", "release_vehicle_to_driver", "get_vehicle_record",
             "complete_safety_inspection", "log_odometer", "hold_vehicle"],
 "before_action": true}

The plan releases the vehicle without the insurance or safety checks it was offered: omission fires and the verdict is deny. This is a good way to try the agent probes. In production, use the API, which checks each tool call as it happens and can hold it for a person.

The API

The API is the part you build on. Create a key under Profile → API keys on neuronlens.com and call POST /v1/score:

  • Answers: send the prompt, what the model read, and the answer. See Score answers.
  • Agent steps: send the run so far and the tool call the agent wants to make, before it runs. See Check agent steps.
  • No code changes: an integration makes the same check for you from Claude Code, an MCP gateway, LiteLLM, OpenTelemetry, LangGraph or Google ADK.

Every call is stored under your workspace and shows up in the console straight away, unless you send "store": false.

The Pro console

The console is the same data, made for people. Open app.neuronlens.com/agent-lens and pick Agent Lens Pro in the sidebar.

ScreenWhat you do there
Get starteda checklist from your first key to your first reviewed flag
Connectyour key, copy-paste code for every way in (with the key filled in), and the apps and agents that have sent traffic
Monitorevery answer and run, a queue of flags to review, held steps to approve, and Why was this flagged?
Evalfailure scenarios, test runs against your agent, and signed audit reports
Settingsprobes, thresholds, the deny level, rules, data retention, alerts, workspaces and team

Older screens (Audit log, Evidence, Probe packs, Activity, Findings and more) are under More at the bottom of the console menu. Some of them are for NeuronLens staff only.

Getting Pro. If you don't see Agent Lens Pro in the sidebar, your account doesn't have it yet. Contact us for access or an invite code.

Moving from the playground to the API

The answer probes are the same in both, so a result you saw in the playground carries over. What changes:

Playground (/v1/agent/score)API (/v1/score)
prompt, optional responseprompt and response, or messages (with tool messages for what the model read)
gpt-oss-20b can write the answeryou always send your model's answer
loci: token weightsspans (the words that drove a probe) and sentences (each sentence's risk)
nothing kepta decision_id, stored and shown in Monitor
answers, and a whole agent run as a log or a plananswers, and each agent step before it runs, with holds and approvals