AI product Open source
three.dev is an AI-agent observability and debugging service from round3ai. It sits as a proxy between an application and providers such as OpenAI, Anthropic, Gemini, and Azure OpenAI, recording production conversations and sending each response to the application and an evaluator for pass/fail scoring.
It attaches evidence to detected issues, groups related issues into failure modes, and ranks them by severity and frequency. Coding agents can read these failure modes through MCP tools such as `list_failure_modes`, `get_failure_mode`, and `get_request_conversation`, then use the recorded evidence to locate and modify the relevant code. The service also compares traffic slices, including before and after a fix or between prompt versions, and reports newly appearing failure modes.
The product can be added through the `three-dev-setup` skill in the `round3ai/skills` repository, which routes an application's LLM calls through the three.dev proxy.
1 use taken from transcripts — each links to the moment in the video.
Detects failures in your AI agent's live traffic and hands fixes to your coding agent, sitting as a proxy between your app and providers like OpenAI, Anthropic, and Gemini, grouping issues into ranked failure modes readable via MCP.
1 in the library.