OpenEik

The open building model for AI agents. The building, stored as a code base.

Opens Q4 2026Open sourceBuilt for LLMs
The era of LLMs is here

IFC won't cut it.

IFC stores what a building looks like, not what it is. A wall arrives as a mesh with its door holes already cut. Ask an AI agent to change it and it edits dead geometry.

  • Move a door, and the old hole stays in the wall.
  • Move a column, and it no longer stacks.
  • Shorten a corner, and the walls no longer meet.
39%

of edits came out right when AI agents edited IFC. Every failing agent reported success.

7 of 18 edits · internal benchmark, October 2026

But

LLMs are amazing at code bases.

They search, read the few files that matter, change one line and run the checks. So we made the building a code base.

One record per file, in meaningful folders. Intent instead of meshes. Bodies, joins, openings and quantities are derived by an engine that checks every change before it is saved.

model/ ├ architecture/level-0/wall/w-3.json ├ architecture/level-0/door/d-12.json ├ structure/level-3/column/c-12-6.json └ types/architecture/wall/yv7.json $ ek request model '{"action":"execute", …}' - "along": 9000, + "along": 14000, ✓ saved · 1 file · the engine re-cut the wall
The same edits, on OpenEik

Right every time. And faster.

100%of edits right on OpenEik
39%on IFC, with the same agents
2.8×faster to a correct edit

Internal benchmark, October 2026: 18 edits per format on two large buildings, made by Claude Sonnet agents and judged by geometry against the engine.

Fast.

An agent reads a model in tenths of a second and changes it with a validated command.

Reliable.

Every write is checked against the whole model. A change that would break it is refused with a reason.

Open source.

The standard and its engine core are open. IFC stays as a bridge for export and import.

The trailer
Coming next

EikHub

Where buildings branch, get reviewed and merge, the way software does.

OpenEik opens in Q4 2026: the specification, the engine core and the benchmark.

github.com/openeik