Pan-Chera/Multi-Agent-CAD
MAC claims text-to-CAD generation at 116× fewer tokens and 13× lower inference cost than a single-agent baseline.
The project splits CAD generation across four agents that pass structured JSON and QA artifacts instead of replaying full context. Its README reports a 99.3% feature pass rate on a 10-prompt, 141-feature benchmark, versus 97.9% for the cited CAD Skills baseline. It includes terminal and web UI workflows for generating STEP/STL outputs, previewing models, and intervening during repair loops. The authors say the system can route stages to different OpenAI-compatible model providers and uses a deterministic translator for common CAD operations. GitHub · LLM repos' note
The project splits CAD generation across four agents that pass structured JSON and QA artifacts instead of replaying full context. Its README reports a 99.3% feature pass rate on a 10-prompt, 141-feature benchmark, versus 97.9% for the cited CAD Skills baseline. It includes terminal and web UI workflows for generating STEP/STL outputs, previewing models, and intervening during repair loops. The authors say the system can route stages to different OpenAI-compatible model providers and uses a deterministic translator for common CAD operations. GitHub · LLM repos' note
score 4