BimClaw

An AI agent that operates on building information models — query, analyze, and modify BIM data through a structured tool surface.

BimClaw is an AI agent for the BIM (Building Information Modeling) workflow. It speaks the language of building structure — walls, doors, columns, slabs, levels — and exposes a tool surface that lets a language model query, analyze, and edit a project the way an architect would think about it.

How it works

  • BimDown format: a text-based BIM file format designed to be AI-native — CSV tables of typed entities plus per-level SVG geometry. The format is the data structure the AI reasons over.
  • Tool surface: bimdown_build (validate + sync), bimdown_schema (introspect tables), bimdown_query (DuckDB SQL over loaded entities), bimdown_render (silent floor-plan render for verification), bimdown_resolve_topology (auto-resolve MEP curve endpoints).
  • Validation loop: every write goes through immediate validation; the agent sees errors and self-corrects.

Why

Conventional BIM software (Revit, ArchiCAD) was built for human-driven workflows and is hostile to programmatic editing. BimClaw treats AI as a first-class user — the data format and the tool surface are designed around what works well for an LLM.