EvalGlass — frequently asked questions

EvalGlass is an AI quality-control tool that lets your coding agent build and run project-specific checks for your AI application. It studies your repo, traces, prompts, and model calls; runs checks locally and in CI; and returns bounded scorecards. See: Overview

No. There is deliberately no certify, approve, or gate verb. A scorecard is a bounded claim about a measured run — it never asserts your AI application is correct, safe, or production-ready. See: No false green

It means EvalGlass measured behavior but no active, host-approved gate ran — so the result is evidence, not enforcement. It exits 0, but exit 0 does not mean quality was enforced. See: Verdicts & exit codes

No. EvalGlass is local-first: the vendored runtime runs on your machine and in your CI, your repo stays the source of truth, there is no telemetry, and no provider keys are required on the run path. See: Local-first

Importing exported traces ships now — local JSONL, OpenTelemetry, and OpenInference exports. A live provider pull (Langfuse, Phoenix, LangSmith) is an experimental, opt-in lane reachable with the connect --live verb — credentials are env-var names, data_policy defaults to unknown, no provider SDK is imported, and a pull is proposed and cannot gate. See: Connectors

Not yet. EvalGlass is v0.1.0 pre-alpha, untagged, and not yet published to PyPI. It is open source under Apache-2.0 and usable today, but contract stability is still settling. See: Roadmap & status

Claude Code is the primary plugin and launch surface. Codex is supported as a second coding-agent runtime through the same evaluation model; its public marketplace listing is a separate next step. See: Codex runtime

You promote it deliberately: validate a dataset, approve a threshold, set the metric to gating, and calibrate a judge if one is used. Authority comes from host-owned records, never from the plugin or the agent. See: Promote a gate