AIBIMTool

AIBIMTool Skills

[License: MIT] [Version] [Skills] [npm] [Website]

AI agent skills for designing buildings as real BIM models with AIBIMTool. Describe a building; an AI architect researches the type, asks the few questions only you can answer, plans large work, then models every storey, wall, slab, door, window, room, stair and roof as IFC4, and changes the model when you ask. Works with Claude, ChatGPT, Claude Code, Cursor, Codex, VS Code, Windsurf and anything else that loads Markdown skills or speaks MCP.

Each user signs in with their own AIBIMTool account, and every message is billed to its credits. The skills drive the public AIBIMTool API through the MCP server, so there is nothing to build and no BIM kernel to install.

[Plan of Level 2 of a 7-storey office the agent built: central core, open workplaces, meeting rooms, private offices and focus booths with their areas] [3D view of the same 7-storey office: a rounded glass bar on a glazed plinth, timber fins, sky lounge and roof terrace]

_{One brief, one IFC4 model of 1,065 elements: the plan of Level 2 with every room measured, and the building itself.}

Install

Pick one. Each method sets the skills up for your agent.

npx skills (recommended, cross-agent)

npx skills add MeltFlexDevs/aibimtool-skills

GitHub CLI (v2.90+)

gh skill install MeltFlexDevs/aibimtool-skills

Claude Code marketplace

Inside Claude Code:

/plugin marketplace add MeltFlexDevs/aibimtool-skills/plugin install aibimtool@aibimtool

The plugin brings the hosted MCP server with it: run /mcp and choose Authenticate once.

MCP server only

If you want the tools without the skills, add the hosted server. It signs you in with your AIBIMTool account and needs no key:

claude mcp add --transport http aibimtool https://aibimtool.com/mcp

In Claude (claude.ai or Desktop) and ChatGPT, add https://aibimtool.com/mcp as a custom connector. For a local server with an API key: claude mcp add aibimtool -- npx -y aibimtool-mcp. Configuration for Cursor, Codex, VS Code and Windsurf is in INSTALL.md and on the MCP page.

Setup script

Universal fallback:

git clone --depth 1 https://github.com/MeltFlexDevs/aibimtool-skills.gitcd aibimtool-skills./setup

More options in INSTALL.md. Agent-driven install, to paste into your agent: INSTALL_FOR_AGENTS.md.

Authenticate

The hosted server at https://aibimtool.com/mcp needs no key: your client opens the AIBIMTool sign-in. Like the API and the CLI, it is part of the Studio plan.

The local server and the CLI use an API key, also part of Studio. The simplest way to get one is the browser flow, which mints a key from your session and hands it back over loopback so you never copy or paste a secret:

npx -y aibimtool-mcp auth login

Or create one by hand at aibimtool.com → Developers → API keys and export it:

export AIBIMTOOL_API_KEY="aibim_sk_..."

The key is shown exactly once at creation. Only its SHA-256 is stored, so nobody can recover it afterwards, including us. An account may hold five live keys and revoking one takes effect on the next request.

Skills

Tools

Metered means a message is charged for the AI it uses: 1 credit per $0.02, at most 150 credits per 8-minute segment. A run that fails on our side is refunded automatically, so an agent that retries does not quietly drain the balance.

A building takes minutes, longer than a chat holds a tool call open, so create_building returns at once and get_project_status waits up to 50 seconds per call while the agent works. The local server writes IFC files and pictures to disk (./aibimtool-output/ unless you pass output_path) and returns the path rather than the bytes: an IFC is megabytes, and the agent usually just hands it to the next step.

What it looks like in practice

You: Design a 7-story office building with a ground-floor lobby, café, and event space. Floors 2 through 4 are typical office floors around a central core with stairs, elevators, and restrooms. Add a rooftop terrace.

[create_building]     brief: "Design a 7-story office building with a ground-floor lobby, …"→ project 019cf06362ab4dac · running
[get_project_status]  wait_seconds: 50→ needs_answer  1. Floors 5-7: Same typical office floors / Offices plus a top-floor amenity / Larger-tenant floors  2. Floor plate: About 800 m² / About 1,200 m² / About 1,800 m²  3. Massing: Rounded bar on a glazed plinth / Two offset volumes with a cantilever / Stepped terraces
> You: Offices plus a top-floor amenity, about 800 m², rounded bar on a glazed plinth.
[send_message]        → needs_approval: a five-task plan> You: Go ahead.[approve_plan]        → running[get_project_status]  → ready · 1,065 elements
[get_model_image]     view: plan · storey: 1→ ./aibimtool-output/019cf06362ab4dac-plan-storey-1-2026-10-28-14-02-09.png[get_ifc_download]→ ./aibimtool-output/design-a-7-story-office-building-with-a-2026-10-28-14-02-11.ifc

That is a real run: the questions, the plan and the 1,065 elements are what the agent produced for that brief. The IFC opens in Revit, ArchiCAD, BlenderBIM, Tekla and Navisworks; the project also opens in the AIBIMTool 3D viewer to keep changing it. More patterns in COOKBOOK.md.

A good brief changes everything

The agent builds what the brief says and researches the rest, stating its assumptions. A brief that names the building type, its size, the number of storeys and the rooms that matter gets a design on the first run; a vague one gets questions first. Pass the user's words, add what they already told you, and never invent requirements to make it look complete. It is the single biggest quality lever in the whole workflow.

The CLI

The same package is a command-line tool when you run it with arguments instead of letting an agent speak MCP to it:

aibimtool build "Kindergarten for 60 children, one storey, 4 groups" --wait --yes --ifc kindergarten.ifcaibimtool send 019cf06362ab4dac "Make the event space open to the terrace" --waitaibimtool image 019cf06362ab4dac --storey 1 -o level-2.png

Full reference at aibimtool.com/developers/cli.

Documentation

• API reference: every endpoint, request and response
• CLI: every command and flag
• MCP: client-by-client setup
• COOKBOOK.md: worked examples
• CONTRIBUTING.md

Licence

MIT. See LICENSE.