Hey Cursor community,
Full disclosure upfront: This is not a sale or an ad—I haven’t even turned on payment processing yet while validating proof-of-concept, so every plan up to Scale is 100% free to try with zero customer data ever leaving your machine.
I’m 15 and spent the last few months building Wallras (https://wallras.com): a single-container, self-hosted transparent proxy (base_url swap, ~3–5ms overhead) that sits between your AI app and OpenAI / Anthropic / Gemini / Ollama / vLLM, paired with a zero-dependency MCP server (wallras-mcp) designed specifically for Cursor:
How It Works With Cursor
- 1-Prompt Agent Setup (
llms.txt): Point Cursor athttps://wallras.comand connect the zero-dependency Python MCP server (wallras-mcp). - Autonomous Debugging & Evals: Cursor can inspect live drift alerts, pull failing traces, diff prompt versions, and run regression datasets right from your editor (read-only by default, with
--enable-writesopt-in). - Gated Prompt Releases: Cursor can only promote a prompt change to
productionif it passes your regression test suite first.
What the Proxy Checks Under the Hood
- Silent Model Swaps & Drift Radar: Uses permutation-calibrated PSI and Jensen-Shannon divergence to catch when an upstream API quietly changes model behavior or spikes in refusals—without spamming false alarms on normal temperature variance.
- Logprob Confidence Mismatch: Flags when a model sounds 95% sure in text while token probabilities were actually guessing.
- 100% Local & Security-Hardened: Runs on your own SQLite or Postgres database with optional AES-256-GCM encryption, tamper-evident hash chains with zero-overwrite deletion, Sigstore-signed Docker images, 366 automated attack/unit tests, and A+ scores on Mozilla Observatory and SSL Labs.
Direct Links
- Homepage & Quickstart: https://wallras.com
- Docs & MCP Setup (
llms.txt): wallras.com
Where I need help: I’m building this on a basic setup without heavy production traffic or big API budgets. I’d really appreciate any feedback from Cursor MCP users on the tool schemas, onboarding flow, or anything that breaks on real workloads!