Launching today
Wallras
Catch AI regressions before your customers do.
1 follower
Catch AI regressions before your customers do.
1 follower
Wallras is one self-hosted container between your app and your AI provider. It catches drift and silent model swaps, traces agent runs, tests prompts before release, flags shaky answers and keeps a tamper-evident log. Your data never leaves your servers.







Free Options
Launch Team

ZethosThe ledger you and your AI work from
Promoted
Hey Product Hunt 👋
I built Wallras after noticing that AI apps often break without any error: the provider quietly updates the model behind the same name, an agent gets stuck in a loop, or a prompt tweak makes answers worse. Nothing crashes, so normal monitoring stays green while users notice first.
Wallras is one container you run between your app and your AI provider (OpenAI-compatible APIs, Anthropic and Gemini). Change one base URL and it watches every request:
• Drift and silent model swaps: length, wording, refusals and the served model version, compared with each model's own baseline. It only flags changes bigger than chance, so you're not chasing noise.
• Agent traces with loops and tool errors flagged
• Regression tests and prompt versions: a failing test blocks the release
• Answers that sound sure but aren't, and topics where your AI is often wrong
• Cost, latency and usage per model and app, with alerts to Slack or PagerDuty
• A tamper-evident log of every answer, optionally encrypted with a key only you hold
• An MCP server, so Claude or Cursor can ask "what's drifting?"
Your prompts and answers never leave your servers. The free Hobby plan includes the drift checks.
Want to see it before installing? Model watch runs the same checks on 7 public models every day: wallras.com/model-watch
How is this different from other AI tools?
• Observability tools show you logs and traces. Wallras also tells you when the model itself changed: it compares each model with its own baseline and only flags differences bigger than chance.
• Testing tools check your code before release. Wallras watches real production traffic, every day, after release.
• It runs entirely on your servers, so prompts and answers never leave you, and every answer is in a tamper-evident log.
• You can see it work before installing: wallras.com/model-watch checks 7 public models daily.
What would you want it to catch next? I read every comment.