What it is
What EvalGlass is — and what it isn't
EvalGlass is an AI quality-control tool that lets your coding agent build and run project-specific checks for your AI application. Knowing what it deliberately is not is part of the product: the honest boundary is what keeps a scorecard worth reading.
It is / it is not
| EvalGlass is | EvalGlass is not |
|---|---|
| A local-first, open-source AI quality-control tool you operate through your coding agent. | A hosted platform, dashboard, or telemetry product. |
| A way to build project-specific checks from your repo, traces, prompts, and model calls. | A generic benchmark suite or a leaderboard. |
| A producer of bounded scorecards with honest verdicts: informational, pass, fail, blocked. | A certification that your AI is correct, safe, or production-ready. |
| Repo-native: your project keeps the truth and the authority to gate. | Something that grants gate/approve/certify authority by itself. |
| Led by the Claude Code plugin, with Codex supported as a second runtime. | Tied to a single coding-agent runtime. |
The ground-truth floor
We state the project's real status plainly wherever it matters. EvalGlass is
v0.1.0 pre-alpha, untagged, and not yet on PyPI. It is
Apache-2.0 and local-first: it runs on your machine and in CI, sends no telemetry,
needs no provider keys on the run path, has no hosted platform,
and ships with empty authority by default — a fresh install gates nothing. There
is no gate, approve, or certify verb.
Read the honesty model in full: Honesty charter · No false green · Local-first · Roadmap & status · Status (docs).