Every AI agent hallucinates. Claude Code does it. ChatGPT does it. Every major model does it.
The problem isn't the model — it's that no one is checking the output before it reaches your workspace.
I spent months watching agents confidently return wrong code, invented API calls, and fake file paths. Then I built a verification layer that catches all of it.
What it does
- 13 detectors that scan every output for hallucinations, safety refusals leaked as content, fabricated citations, system prompt leaks
- 31 correction strategies that fix issues automatically
- Knowledge graph cross-referencing to validate factual claims
- Model-agnostic — works with Claude, GPT, DeepSeek, Llama, any provider
- 0 GPU required — runs on CPU in under 100ms
Why I built it
I was losing hours to hallucinated outputs. An agent would confidently tell me it had edited a file — but the file was unchanged. It would fabricate API responses that looked real but didn't exist.
The verification layer sits between the model and your workspace. It doesn't just flag issues — it surfaces a correction before the output touches your codebase.
How to get it
Download: https://agent-download-site.vercel.app
Free, model-agnostic, CPU-only. No strings attached.
Built as an open-source tool for the AI community.
Top comments (0)