See detailed traces and spans, tool calls included
Every LLM call, tool call, retrieval and database query, nested parent to child on a timing waterfall.
See detailed traces and spans, tool calls included: Docs (opens in new tab)Observability
Every agent decision,
captured. Not just a mere log.
Add the SDK and every reasoning step, tool call, retrieval and handoff your agent makes in production becomes a structured, replayable trace.
Trusted by teams shipping agents in production
What is Observability?
A structured record of every step your agent takes. Netra captures each LLM call, tool call and retrieval as a span inside a trace, with inputs, outputs, latency, tokens and cost. Sessions, users and tenants group those traces, so you can go from a customer's complaint to the exact step that caused it.
Find the step that failed, not just the request that did
Every LLM call, tool call and retrieval is its own span on a timeline, with inputs, outputs, latency and tokens. A bad answer points at the step that caused it.
See the metadata of every step
Inputs, outputs, parameters and type-specific details such as model names and token counts, one click into any span.
See the metadata of every step: Docs (opens in new tab)
See cost, latency and tokens on every span
Spot the slow or expensive step, then roll the numbers up by model, tenant, user or time range.
See cost, latency and tokens on every span: Docs (opens in new tab)
Tell every step apart by type
Span, generation, embedding, tool and agent types are assigned automatically, so each step shows up as what it is.
Tell every step apart by type: Docs (opens in new tab)Know what every customer costs you
Set a tenant ID once in middleware and every trace, token and dollar after it is attributed to that customer. The same views work for every agent, session and user.
Switch between agent, tenant, session and user views
Agent cost, latency and tool use; tenant spend; sessions as whole conversations; every trace for one user.
Switch between agent, tenant, session and user views: Docs (opens in new tab)Attribute cost and usage to each tenant
Sessions, traces and total cost for every customer, sorted by spend, so the heaviest users are one glance away.
Attribute cost and usage to each tenant: Docs (opens in new tab)Debug the journey, not 40 requests
Group multi-turn interactions by session and user, then drill from session totals to the exact span.
Debug the journey, not 40 requests: Docs (opens in new tab)Save the views your team checks
Search by name, time range, status or any custom attribute, choose your columns, and save the filter.
Save the views your team checks: Docs (opens in new tab)See how the whole system is behaving
Latency percentiles, error rates, cost and tokens across all of your traffic, out of the box. Then build the dashboards your team will actually open.
See detailed analytics of your traces
Latency percentiles, error rate, request volume, cost, tokens and time to first token across every agent.
See detailed analytics of your traces: Docs (opens in new tab)Build custom dashboards
Six chart types over the same metrics, segmented by environment, user, session or tenant.
Build custom dashboards: Docs (opens in new tab)Filter down to what matters
Narrow by model, user, session, error or PII flags, and by latency or cost range.
Filter down to what matters: Docs (opens in new tab)Reuse widgets across dashboards
Add a widget to as many dashboards as you like, or clone a whole one, so every team reports on the same numbers.
Reuse widgets across dashboards: Docs (opens in new tab)Fix it where you work, and make sure it stays fixed
Pull the failing trace into your IDE through MCP, then turn it into a test case. The same bug can't ship twice.
Debug using MCP
Bring traces, spans and sessions into Cursor, Claude Code, Windsurf or Codex through the Netra MCP server.
Debug using MCP: Docs (opens in new tab)Look up any trace or session
Fetch spans by trace ID, filter traces with pagination, and inspect session metrics without leaving your editor.
Look up any trace or session: Docs (opens in new tab)Add a trace as a test case
Turn a failing production trace into a test case, so the bug you just found becomes a check in your next evaluation.
Add a trace as a test case: Docs (opens in new tab)Grow a dataset from your editor
Create a dataset and add items one at a time or in bulk, straight from your coding agent.
Grow a dataset from your editor: Docs (opens in new tab)Keep your stack, add the context it's missing
Built on OpenTelemetry with custom exporters, so Netra adds agent-level context without asking you to rip out what already works.
- OpenTelemetry
- LangChain
- LangGraph
- CrewAI
- LlamaIndex
- OpenAI
- Anthropic
- LiveKit
- Vapi
- Pinecone
- Datadog
- MCP
Instrument once, in minutes
One Netra.init() traces 30+ AI libraries, including model providers, frameworks and vector databases, with no code changes.
Add decorators and manual spans
@workflow, @agent, @task and @span for minimal-code tracing, plus manual spans for anything custom.
Trace voice agents end to end
Speech-to-text, the reasoning LLM and text-to-speech land in one trace, so a delay shows up on the stage that caused it.
Docs (opens in new tab)Control what's captured
Choose which libraries are instrumented, and switch off prompt and response capture while keeping latency, token and cost metrics.
Docs (opens in new tab)In production
Every agent, every tenant, in real time
How Pencil watches every agent, every tenant
Netra turned our ad-agent stack from a black box into a clear control center. We can see what every agent is doing, for every tenant, in real time, and Netra now processes over 1M spans every day for Pencil. Issues that once took hours to debug now take minutes.
Frequently Asked Questions
Everything You Need to Know About Observability
Can't find the answer here? The Observability docs go deeper, or talk to our team.
What does a Netra trace include?
The complete execution of a request: LLM calls with prompts, completions and model parameters, tool invocations, retrieval steps, token usage, latency per step, cost, errors and any custom metadata you attach.
What's the difference between a trace and a span?
A trace is the whole journey of one request. Spans are the individual operations inside it — each LLM call, tool call or retrieval — arranged parent-to-child.
Which ways can I instrument my agent?
Three, and they mix: auto-instrumentation with no code changes, decorators such as @workflow, @agent and @task, and manual spans for full control.
How do I track agents, sessions and users?
Use the @agent decorator for agent execution, set_session_id to group related interactions, and set_user_id to attribute traces to a person. set_tenant_id does the same for a customer.
How are costs tracked?
Token usage and cost are captured automatically for supported providers and can be broken down by model, tenant, user or time range. For a custom gateway or unsupported provider, attach usage as span attributes.
Can I keep prompts and responses out of Netra?
Yes. Content capture can be switched off while latency, token and cost metrics are kept, and you choose exactly which libraries are instrumented.
Start today
See the step that broke before the ticket lands
Install the SDK and see your first trace in minutes. Free to start.