Skip to main content
Vizra Cloud is the hosted counterpart to vizra/evals. Same engine, same tables, same scores — it just keeps them somewhere your whole team can see, across every environment your evals run in.
The Vizra Cloud overview: suites needing attention, then recent runs across every project

​
Local recording happens either way

This is the part worth being precise about, because it decides whether you need Cloud at all: Nothing about installing vizra/evals sends us anything. Runs persist to your tables and the local dashboard reads them, key or no key. Cloud is additive: set VIZRA_CLOUD_KEY and finished runs are also pushed to your project.
Nothing executes on our side. Your evals run in your app, against your models, with your keys. Cloud stores what your runs produced — which is why it never becomes a runtime dependency of your test suite, and why an outage on our end cannot break your build.

​
What the key actually buys

  • One history across every environment. A run from your laptop, a run from CI and a run someone triggered from the browser all land in the same suite, tagged with the environment they came from — so “does this pass on main” and “does this pass on my branch” stop being the same number.
  • Your team’s runs, not just yours. The local dashboard shows what your machine did. Cloud shows what everyone’s did.
  • Triage first. The overview leads with suites that fell below their baseline or broke outright, worst first, across every project — rather than leaving you to click into each one and read the score.
  • CI context. Runs carry the build URL, branch and pull-request number, detected automatically from GitHub, GitLab, CircleCI or Buildkite.
  • Dataset editing. Change a test question in the browser, run the variant, and export it back to your repository — without booking a developer’s afternoon.
  • Running a suite from the dashboard, on your own infrastructure.

​
Every tier records

Runs are never metered and reporting is never rate-limited — on any plan, including the free one. The plans differ on projects, seats and how long per-sample detail is kept, not on whether your evals are allowed to report.

Connect a project

Project, key, .env — and your next run appears.

What gets reported

Which surfaces push, what’s in the payload, and how to send less.

Editing datasets

Fix a test question in the browser, commit it back to the repo.

Plans & retention

Limits, and exactly what ages out.