spanview

// observability + auto-diagnosis for AI agents

Your agent is broken.
You don't know why.

spanview ingests OpenTelemetry traces from any agent framework, runs an LLM evaluator over every step, and hands you a plain-English verdict — which step failed, why, and what changed since the run that worked.

Get early access See how it works

works with LangGraph·OpenAI Agents SDK·Claude Agent SDK

run · research-agent 6 spans
  1. 01AGENTresearch-assistant2.6s
  2. 02LLMplan840ms
  3. 03TOOLsearch120ms
  4. 04LLManswer1.1s
error trace_report

Stale tool result at step 3

The search tool returned a cached hit from a prior query; the agent answered from it without re-checking, so the final response is wrong.

  • step 02 · LLM plausible plan, no validation
  • step 03 · TOOL stale cache hit

// the problem

Agents fail quietly. You find out from a user.

01

Silent failures

A tool returns wrong data. The model invents a step. The output looks plausible and ships. Nothing throws.

02

Unreadable traces

Raw OTel spans tell you what happened, not why it's wrong. Forty spans a run, read by hand, doesn't scale.

03

Invisible regressions

You change a prompt, a tool, a model. Something breaks. Which change? Which step degraded? No idea.

// how it works

Instrument once. Read the verdict.

  1. 01

    Instrument once

    One auto-instrumentor. No code changes, no manual spans.

    from openinference.instrumentation.langchain import LangChainInstrumentor
LangChainInstrumentor().instrument()
  2. 02

    Traces flow in

    Spans land over OTLP in real time. spanview normalizes OpenInference, OTel GenAI, and OpenLLMetry — no adapter.

  3. 03

    Evaluator reads every step

    An LLM walks the trace and leaves a note on each span: what it expected, what it got, whether the step is sound.

  4. 04

    Verdict & diff

    Open the UI, CLI, or MCP — the diagnosis is already there. Diff two runs to pin the regression.

// what you get

Everything points at the diagnosis.

Per-step diagnosis

Every span gets an LLM note — not pass/fail, a sentence on what happened and why it matters. Severity-ranked.

Run comparison

Diff any two traces: inputs, outputs, tool calls, tokens, step order. Regressions highlighted automatically.

Any framework

OpenInference, OTel GenAI, OpenLLMetry — all normalized into one model. Bring whatever stack you run.

UI · CLI · MCP

Read verdicts in the browser, from the terminal, or via an MCP server plugged straight into Claude.

Open source

Self-host on your machine or your infra. Traces never leave your environment. Run it for free, forever.

Secret redaction

Built-in patterns strip API keys, tokens, and credentials from spans before anything is written to disk.