Integration · OpenInference
Evaluate OpenInference traces in EvalGlass.
OpenInference is an open tracing convention for LLM and agent calls. You bring traces you have already exported; EvalGlass normalizes them locally into its vendor-neutral trace model and your coding agent turns them into project-specific checks. Exported import ships now. A live provider pull is an experimental, opt-in lane (the connect --live verb) — nothing is phoned home and no SDK runs on the evaluation path.
Import my exported OpenInference traces and build checks around them.
What ships today
Import of exported OpenInference traces (file/JSON) → local normalization to a TraceEnvelope → project-specific checks → a bounded scorecard. No vendor dashboard owns the result; your repo stays the source of truth.
-
01 · your traces
Your tracesYour app emits OpenInference spans — the OpenTelemetry convention for LLM and agent traces.
-
02 · export → import
Export, then importYou export them to a file (JSON). EvalGlass reads it locally and maps it to a vendor-neutral TraceEnvelope — no SDK, no network, no keys.
-
03 · evaluate
Evaluate locallyYour project-specific checks run on the normalized behavior, entirely on your machine — and a bounded scorecard comes back.
connect --live). Nothing is phoned home and no SDK runs on the evaluation path. Working with traces →The live lane experimental
A live pull directly from OpenInference is an experimental, opt-in lane reachable with the connect --live verb — until then, export from OpenInference and import the file. We mark the difference plainly so nothing reads as more finished than it is.