Archicad AI Pilot— MEPbridge ACAIstr | Version v0.1.4 Release, Read/Write & Automation Validation

[Bleufeng]

Archicad is a genuine forward‑thinking 3D design software natively developed for architects. Cross‑disciplinary integration between architecture, MEP and other disciplines remains a long‑standing bottleneck for industry‑wide adoption.

This project was born from a persistent industry pain point: cross-disciplinary coordination between architectural and MEP design on Archicad has long failed to gain widespread adoption across all teams. With the rise of artificial intelligence, Blender, Rhino and Revit have each rolled out MCP-powered automation capabilities, which motivated me to build this tool as a spare-time side project. While it cannot yet fully support end-to-end real-world construction projects, it has proven a critical breakthrough: AI can read all parameters stored within design files and trigger corresponding software operations. The full technical potential of this solution remains to be unlocked, hinging on the iterative advancement of large language models and future functional expansions. Its core purpose is simply to offer architects an open, experimental technical pathway to explore AI-augmented design. Beyond being a standalone Archicad MCP connector, this tool integrates dual operation modes — AI-driven smart workflow and manual command control — alongside an extensible platform for custom user expansions. The project MEPbridge ACAIstr is open-sourced and available for download on GitHub at https://github.com/bleufeng/mepbridge-acaistr 

所用操作系统：Windows

• General building elements: walls, columns, beams, slabs, roofs, doors, windows, stairs, objects, lamps, meshes, zones, and grids
• MEP domains: pipe, duct, cable carrier, flexible segment, take-off, systems, sizes, and route operations
• Model editing: move, rotate, mirror, copy, geometry changes, batch creation, selection, and deletion
• Project queries: project information, stories, libraries, hotlinks, properties, geometry, mesh, and viewport capture
• CollabAI Smart and Manual operation modes with risk classification, confirmation gates, and result readback
• MCP tools generated dynamically from ai-adapter/tool-descriptors.json
• Local user templates, custom commands, knowledge base, learning memory, and audit logs
• Reviewed Workbench modules loaded from modules/registry.json
• AC28/AC29 support with preview, confirmation, result readback, and failure handling according to operation risk

[Bleufeng]

v0.1.2 -Changed
Installation guides document the Node.js prerequisite, custom install paths, and a full step-by-step manual installation procedure covering administrator rights, locating the Archicad folder, choosing between Archicad 28 and 29, handling a previous build, and verifying the result.
The registered command and descriptor counts now come from a single ai-adapter/command-boundary.json file that the release gate, the public repository validator, CI, and the packaging scripts all read, instead of each carrying its own copy.
Command examples no longer contain hardcoded MEP system names or attribute indexes. Those identifiers are not stable across projects, locales, or Archicad versions, so omitting them lets the Add-On use the domain default. A static check keeps them out.
The Archicad runtime regression now checks returned content rather than only that response fields exist, and covers the favourite, layer-batch and classification write commands end to end, each restoring what it changed.
The internal command guide lists all 74 registered commands. It had 52, so the favourites, profile, view-map, classification and layer commands were absent along with the project and story queries. Command counts there now defer to ai-adapter/command-boundary.json.
The public repository validator no longer aborts on a Markdown link containing a stray percent sign; such a link is reported as broken instead.

[catalinpopa]

claude is also working on this, what s the difference?

[Bleufeng]

This refers to Claude Design?  While Claude, GPT and Gemini already have their respective native MCP connections, they lack professional parsing for Archicad’s internal models. From an architectural‑specialty perspective, our goal is to add a dedicated BIM‑MCP vertical‑adaptation layer for Archicad, translate complex BIM model data into structured data readable by any large language model, and provide open, customizable extensions for any interface.

[Bleufeng]

v0.1.3 -Changed

Fixed

• Slab, Roof, and Zone creation now accepts closed polygon contours.
• Zone creation now uses the supplied contour and correctly writes the zone name
  and category.
• Explicit slab thickness is now applied correctly.
• Degenerate polygon input now returns a clear validation error.
• Copying Mesh and Morph elements is now supported.
• MCP client status now uses the initialize handshake and distinguishes
  connected clients from configured clients.
• Public CI now exports only the approved command-safety test from the private
  test directory.

Added

• Added a read-only endpoint for listing running Archicad instances and their
  open projects.

[Bleufeng]

v0.1.4 -Changed

Fixed

- Explicit wall heights are no longer clamped to the story height.
- Walls now use the center reference line; the body no longer shifts sideways.
- Window sill height is applied; openings above the host wall top are rejected.
- Windows and doors inherit the host wall's story.
- `set_active_story` previews by default and reports `previousStoryIndex`.
- Stair tread depth is reported as the effective `resolvedGoing`.
- Workbench template replay refreshes saved templates before execution.
- Fixed bug causing slow reproduction of user‑defined templates in Workbench.

Added

- Moving and editing building elements by GUID with per-field masks and readback.
- Optional absolute-Z placement for walls, columns, objects, and lamps.
- Mirror and rotate accept explicit GUID lists.
- Four local server tools: list Archicad instances, capture/compare/replay snapshots.
- Copy is all-or-nothing with a single undo; delete is capped at 200 per call with `acknowledgeCount` above 50.