# Wallras

> Wallras is a self-hosted proxy for AI applications that watches every model call and catches what breaks: silent model changes, drift, wrong answers, failing agents and risky prompt changes. It runs as one container on your own servers and works with OpenAI-compatible APIs, Anthropic and Google Gemini. Prompts and answers never leave your infrastructure.

Point your AI client at Wallras instead of your provider (one line), and it records each request, learns each model's normal behaviour, and adds drift detection, agent traces, regression tests, prompt versions with gated releases, a human review queue, gateway controls and a tamper-evident log.

Current engine version: 0.19.0 (install ghcr.io/wallras/wallras-engine:0.19). These files are rebuilt from the site and the engine on every release.

## Pages

- [Wallras: catch AI regressions before your customers do](index.md): Overview, features, pricing and FAQ
- [Docs | Wallras](docs.md): Setup, every feature, configuration and all endpoints
- [How Wallras measures this | Wallras](methodology.md): How each number is measured, tested providers, benchmarks, limits
- [Security | Wallras](security.md): Security model, testing and what is in place
- [Live demo | Wallras](demo.md): A staged demo of ten problems and what Wallras reported

## For agents

- [Full site as one Markdown file](llms-full.txt): every page above
- [OpenAPI spec](openapi.json): every engine endpoint, generated from the engine's code
- [MCP server](assets/wallras_mcp.py): lets an AI assistant check drift, read requests and traces, run regression sets, fetch prompts and work the review queue on your Wallras (read-only unless you turn writes on). Setup: docs.md#agents
- [Python SDK](assets/wallras_sdk.py): report self-hosted model signals, agent steps, fetch prompts
- [Pricing as JSON](assets/pricing.json)
- [Demo results as JSON](assets/demo-results.json): what the engine reported in each staged scenario

## Optional

- [Privacy policy | Wallras](privacy.md): Privacy policy
- [Terms | Wallras](terms.md): Terms of service
- [Sample compliance report](sample-report.html): made by the engine from staged traffic
- [Security contact](.well-known/security.txt)