Archicad is Graphisoft's BIM modeling software, one of the most widely used by architecture firms across Europe for designing buildings in 3D and producing regulatory deliverables. Getting an AI assistant to talk to an Archicad model used to mean scripting directly against Graphisoft's API or doing everything by hand in the interface. The archicad-mcp server changes that: it connects Claude to the model open in Archicad through the Model Context Protocol, Anthropic's open standard for how an AI assistant calls external tools.
The project is published by independent developer lgradisar under the MIT license — neither Graphisoft nor ENZYME-APD are directly involved. It acts as a bridge: it doesn't reimplement any Archicad functionality itself, relying entirely on a second third-party project, the Tapir add-on.
How the technical chain works
The chain has three distinct links, each maintained by a different team:
- Tapir, an Archicad add-on maintained by the ENZYME-APD organization, installs into Archicad and registers roughly 233 documented JSON commands on top of Graphisoft's native ones — element creation, zone management, BCF collaboration, and more. That figure was checked directly against the repo's command definitions file, not a marketing page.
- archicad-mcp connects to Tapir and automatically compiles each of those JSON commands into an MCP tool, with no manual re-coding. This is the server that exposes the tools to Claude.
- Claude Desktop, configured to point at the local server, can then read and edit the open Archicad model within the same working session.
Custom commands can be added on the server side for firm-specific needs, but that means building a native Archicad add-on, not just configuration.
Prerequisites and installation
Before connecting Claude to Archicad, four things need to be in place: an Archicad license on a version Tapir supports (25 through 29, on Windows or macOS), the Tapir add-on installed and enabled in Archicad, a Python environment managed with the `uv` tool, and Claude Desktop installed locally. Setup happens in three steps: clone the archicad-mcp repo, create the virtual environment with `uv`, then edit Claude Desktop's config file to add the server's local path. There's no graphical installer today — this is a setup for a technical workstation, not for someone unfamiliar with a terminal.
What it actually lets you do
Once connected, Claude can drive most routine modeling operations through natural language instead of repeated clicking:
- Building element creation: walls, slabs, columns, beams, roofs, stairs, doors, windows, openings
- Zones and classification: creating zones, zone categories, classification systems and items
- BCF collaboration: creating issues, adding comments, attaching elements to an issue — useful in multi-trade BIM coordination
- Layout work: creating views, layouts, layout subsets and quantity worksheets
- MEP elements: creating and connecting HVAC, plumbing and electrical elements and systems
- Favorites and properties: creating and applying favorites, property definitions and groups
Concrete use case
At the end of a BIM coordination meeting, a project lead can ask Claude to create a BCF issue for every clash detected on a given level, automatically attaching the relevant elements — instead of entering them one by one in Archicad. Another common case in practice: bulk-reclassifying a set of zones after a program change, or generating quantity worksheets for a given trade without manually navigating every view in the project.
Data privacy and security
The server runs locally, on the user's machine: it doesn't send the Archicad project file to any third-party cloud service. Claude Desktop talks to archicad-mcp on the local machine, which in turn talks to Tapir inside the open Archicad instance. The only data that leaves the machine is whatever the user explicitly chooses to send Claude in the conversation, not the full project file. That said, it's still unaudited third-party software: firms under strict confidentiality clauses with their clients should clear it with IT before any production rollout.
Against other BIM MCPs: Archicad vs. Revit
Revit MCP follows an equivalent logic — letting Claude create and modify elements in a BIM model — but with a structural difference: its C# add-in is developed and versioned in the same repo as the server, by the same team. For Archicad, the bridge necessarily runs through a third-party community add-on, Tapir, maintained by a different organization than the MCP server itself. That means the Archicad chain's reliability depends on two independent maintainers staying compatible over time, rather than one unified project.
Cost
Both archicad-mcp and Tapir are free and open source, under the MIT license. The only real cost is one a well-equipped firm already carries: a valid Archicad license on a version Tapir supports.
Who this MCP is (and isn't) for
It makes sense for a firm that already works in Archicad, comfortable with a one-time technical setup (terminal, config file), and looking to speed up repetitive modeling or BIM coordination tasks. It makes less sense for a firm with no in-house technical skills or IT support available for setup, or for a use case that needs a vendor support guarantee — support here comes from the open-source community, not an SLA.
Limitations and caveats
This isn't a Graphisoft or ENZYME-APD project on the MCP-server side: archicad-mcp is maintained by an independent developer, with a still-short commit history (5 commits on the main branch as of this check) and modest adoption (23 stars, 6 forks). Setup remains manual, with no turnkey installer, and depends entirely on Tapir being present and up to date, which only covers Archicad 25 through 29 on Windows and macOS — earlier versions and Linux aren't supported.
Educasium Reading
Key point: Archicad MCP is a useful teaching case for BIM automation when it is treated as a local chain that must be verified step by step. The value is not only the number of Tapir commands exposed. It is the visible path from a natural-language request to a structured action inside a BIM model.
What to teach
A good exercise starts with a test model, not a client project. The learner asks for a limited action: create a zone, list elements, add a BCF issue, or produce a quantity worksheet. Then the result is checked inside Archicad, properties are reviewed, undo history is tested, and a copy is saved.
That method avoids two common misunderstandings. First, the assistant does not magically understand the whole building. It calls available commands, receives responses, and acts inside Tapir's limits. Second, a technically successful command can still be poor BIM work if classification, level, property, or naming does not follow the firm's conventions.
Firm governance
For an architecture team, serious deployment needs a small protocol: exact Archicad version, Tapir version, archicad-mcp version, Claude configuration file, workstation permissions, test folder, save-before-action habit, authorized command list, and rollback procedure. This is not paperwork for its own sake. It separates AI experimentation from traceable professional production.
Confidentiality is the other governance point. The server runs locally, but the assistant conversation can still contain project information. A firm should define what may be described in prompts, what must be anonymized, and which projects are excluded from trials. Competitions, confidential files, or sensitive public projects need approval before automation touches them.
Operational comparison
Compared with Revit MCP, Archicad MCP has a more distributed architecture: Tapir on one side, the MCP server on the other. That split can work well if both projects remain maintained, but it requires compatibility checks. For Educasium, this is a strong lesson in client lifecycle: install, test, document, train, monitor updates, and decide whether the time saved justifies maintenance.
The right question is not "can Claude modify Archicad?" It is "can the team verify what Claude modified?" When the answer is yes, the connector becomes credible for learning BIM coordination, repetitive task automation, and domain API thinking. When the answer is no, it remains an interesting demo, but not yet a production method.
This also makes Archicad MCP a useful comparison point for client lifecycle. A firm can begin with a low-risk inventory task, move to a controlled modification, then decide whether templates, naming rules, BCF workflows, and quantity worksheets are stable enough for repeat use. The practical reading is therefore a deployment method, not only a connector profile.
Going further
Firms that want to structure their use of Claude beyond this one connector can look at the Mastering AI in Architecture course or the ready-made skills in the Skills hub for architects, covering site visit reports, quantity take-offs and quotes.