See every AI agent. Govern every action.
Free Scanner

Your agents are failing silently. Find out where.

ghostspan, the free scanner from MeshAI Labs, turns last week's trace export into a Silent Failure Report: a readable list of retry loops, degraded tool calls, and cost anomalies that returned 200 OK and shipped a wrong answer anyway.

How It Works

Three steps, no upload, no instrumentation if you already emit OTel.

01

Export your traces

Works with OpenTelemetry trace exports (OTLP JSON) and Langfuse exports. Nothing to instrument if you already emit OTel.

02

Run the scanner locally

One pipx install. Your traces never leave your machine: no upload, no API key, no vendor account.

03

Read the report

Counts and concrete examples of what failed silently: retry loops, tool-call success degradation, latency and cost anomalies, quality-drop signals.

pipx install ghostspan

Why Local-First

Trace exports contain your prompts, your customers' data, and your system internals. A diagnostic that asks you to upload them is asking too much. ghostspan runs entirely on your machine and the report is yours. If it finds something and you want a second pair of eyes, that part is up to you.

Why This Exists

Agent failures rarely throw. The HTTP status is 200, the trace is green, and the answer is wrong. Per-trace pricing punishes teams for looking at 100% of production, so most teams sample, and silent failures live in the part they did not sample. A scanner over the export you already have costs nothing to run and has no incentive to meter you.

Get a sample report

One email. No sequence, no SDR follow-up. We will ask one question: did the report find something real?