ARCHIPRACTICA

Independent publication. Every tool tested on the same real brief. No sponsored reviews.

G–05Guide

How to Use Claude AI in Revit: MCP Setup, Plugins, Limits

On this page
  1. What can Claude actually do inside Revit?
  2. Which route should you pick? Four ways to connect, compared
  3. How do you set up the official Revit MCP Server with Claude Desktop?
  4. Our two prompt tests: what Claude got right, and what we could not test
  5. Vibe coding: Claude writes your pyRevit scripts
  6. What does it cost?
  7. Four ways this breaks in a real office
  8. FAQ
  9. Sources

TL;DR: If you are asking how to use Claude AI in Revit, there are four working routes as of October 2026: Autodesk’s official Revit MCP Server add-on (free, Revit 2027.2 only, Technical Preview), community MCP servers for Revit 2020-2027, commercial plugins such as NonicaTab and Geopogo, and no-live-link workflows where Claude writes pyRevit scripts or reads exported IFC files and schedules. Whichever route you take, start with read-only access on a copy of the model, and treat every AI edit and every generated script as a draft that a human reviews before it touches project work.

This guide covers routes for your Revit version, official server setup, two tested prompts, and common failures. The recommendations are deliberately conservative: a write-capable server connected to a live project file is the main risk I see.

This guide is Windows-centric, because every live connection requires Claude Desktop or Claude Code on the same Windows machine as Revit; per techdrawai’s October 2026 review (8 October 2026), Mac users need a Windows VM, and Autodesk documents Parallels. Prices and versions were checked on 20 October 2026. The guide does not cover Autodesk’s own Assistant in depth, and it leaves out ChatGPT-specific plugins. Our pillar guide Claude for architects covers chat-only workflows and promised this Revit follow-up.

What can Claude actually do inside Revit?

Through an MCP server, Claude Desktop can query a live Revit model in plain language, read element data, select and zoom to elements, open views, export views and schedules, and with write-tier servers edit parameters and move elements. MCP is an open standard for wiring AI assistants into other software, “a USB standard for AI” in Autodesk’s June 2026 announcement. There is no Claude panel inside Revit; Claude calls the tools the server exposes.

For everyday use, see the example prompts in Autodesk’s Revit MCP Server documentation (read 20 October 2026). Autodesk uses “List all doors on Level 1”, “Show the fire rating for all doors” and “Create a dashboard showing the number of walls, doors, and windows in the model”. These model-reading questions show how to use Claude AI in Revit right now: query the model, verify the answers, and leave writes for later.

Browser Claude and mobile Claude cannot reach Revit at all as of October 2026; Anthropic’s connector directory had no Revit entry as of 8 October 2026 (techdrawai checked), and meltflexai’s 18 October 2026 guide confirms the same gap. Every working route in this article therefore runs through Claude Desktop or Claude Code on the Revit machine.

On AI adoption across the profession, the RIBA AI Report 2025 (survey January-April 2025, published June 2025) found 59% of UK practices using AI at least occasionally, up from 41% in 2024. RIBA notes the sample is self-selecting.

Which route should you pick? Four ways to connect, compared

On Revit 2027.2, pick the official Autodesk server: free, read-only to start, no arbitrary code. On Revit 2020-2026, pick a community MCP server if someone in the office reads READMEs critically, or a commercial plugin if nobody does. On any Revit version, you can skip the live link and let Claude write pyRevit scripts or read IFC and schedule exports today.

CriterionOfficial Revit MCP ServerCommunity MCP serversCommercial plugins (NonicaTab, Geopogo)No live link (scripts + exports)
Revit versions2027.2 only (Technical Preview)2020-2027, varies by project2020-2027 (NonicaTab); 2025-2027 (Geopogo)any version
Can change the modelRead: no; Write/Experimental: yesyes, most run AI-written codeNonicaTab FREE: no; PRO and Geopogo: yesonly through scripts you run yourself
Arbitrary code executionnooften yes (C# or Python)no (fixed tool lists)you review the script before running
Network exposurelocal stdio process (no network port)varies: port 8080 on all interfaces with no auth (most-starred) to localhost+tokenGeopogo: local only (vendor states); NonicaTab: not statednone
Costfree with a Revit 2027 licencefree, open sourceNonicaTab FREE 0 / PRO 85 EUR + VAT per user/year; Geopogo $19.99/monthfree
Setupabout 20 minutes30-60 minutesinstaller + toolbar buttonnone beyond pyRevit
Who it’s forfirms already on 2027.2offices with a technical personteams that want two-click setupeveryone, today

Feature comparison from vendor documentation and project READMEs, 20 October 2026; versions: official server Technical Preview on Revit 2027.2, star counts and prices as of that date. We did not run these four routes side by side; the setup-time estimates are meltflexai’s, from their published setups (18 October 2026).

Autodesk released the official server on 17 June 2026 as a Technical Preview. On Revit 2027.2 it ships as three separate installers, Read-Tools, Write-Tools and Experimental-Tools, downloaded from Autodesk Account under the Revit 2027 entitlements. The Read tier exposes 8 tools; the Write tier adds parameter edits, moves and rotations; Experimental adds sheets, schedules, rooms, view creation, tagging and IFC/gbXML export (counted from Autodesk’s tool reference, read 20 October 2026). None of the official tools runs arbitrary code.

The original community revit-mcp project is archived, and its continuation mcp-servers-for-revit (389 GitHub stars on 20 October 2026, MIT, Revit 2020-2026) can create, modify and delete elements and can compile and run C# that Claude writes. Per techdrawai’s code read (8 October 2026), that server listens on TCP port 8080 on every network interface with no authentication.

The pyRevit route mcp-server-for-revit-python (214 stars) is better guarded: it binds to loopback only (reachable only from the same PC), pairs with the Windows user, and shows model-changing payloads for review inside Revit before running them. bimwright/rvt-mcp (92 stars, Apache-2.0, Revit 2022-2027) adds token auth (a shared secret the client must present) and a read-only mode; meltflexai specifically recommends rvt-mcp with —read-only for Revit 2025 and 2026.

NonicaTab FREE gives a read-only AI Connector with 36 tools; NonicaTab PRO (85 EUR + VAT per user per year) unlocks the full connector with 50+ tools, and both cover Revit 2020-2027 (nonica.io, 20 October 2026). Geopogo’s AI Workflow Connector costs $19.99 per month on the Autodesk Marketplace (20 October 2026). Geopogo runs a local HTTP server on port 8765, wraps writes in undoable Revit Transactions, and states that no model data leaves the machine.

How do you set up the official Revit MCP Server with Claude Desktop?

Install the add-on from Autodesk Account, add one JSON entry to claude_desktop_config.json, restart both programs, and verify with one prompt. You need Revit 2027.2 (the server will not run on older builds), the Read-Tools add-on, and Claude Desktop on the same Windows machine. Budget about 20 minutes. Work on a copy of the model for the first session.

  1. Update Revit to 2027.2.

  2. In Autodesk Account (accounts.autodesk.com), find the Revit MCP Server Read-Tools Technical Preview under the Revit 2027 entitlements and run the installer. Start with Read-Tools; add Write later if a project needs it.

  3. In Claude Desktop, open Settings, then Developer, then Edit Config. That opens claude_desktop_config.json; on Windows it lives under %APPDATA%\Claude\ per the MCP project’s Claude Desktop guide (read 20 October 2026), and Anthropic’s Help Center covers Developer settings and the one-click extension route.

  4. Add the server entry and save.

    {
      "mcpServers": {
        "Revit MCP Read-Tools Technical Preview": {
          "command": "C:\\Program Files\\Autodesk\\Revit 2027 MCP Server Read-Tools Technical Preview\\Autodesk.RevitMcpServer.Stdio.exe",
          "args": ["--ManageConfigurationWithAutodeskRevitPluginInstaller=true"]
        }
      }
    }

    The entry comes from Autodesk’s setup page (read 20 October 2026). Every backslash is doubled; Autodesk’s troubleshooting page says a single-backslash path fails even when it is otherwise correct.

  5. Quit Claude Desktop completely and reopen it. Start Revit 2027.2 and open a model.

  6. Type the verification prompt Autodesk suggests: “List the available Revit MCP tools.” A returned tool list means the connection works. Claude Desktop asks for your approval before each tool call.

The known limitations in the Technical Preview, per Autodesk’s troubleshooting page (read 20 October 2026):

  • Worksets are not supported.
  • Elements in linked Revit models are not supported.
  • The server does not respond while Revit is in an editor mode or a blocking dialog is open; the Modify tool is the exception.

In the Tech Preview launch thread on the Autodesk forum, an Autodesk-side technical reply (read 20 October 2026 via search extract; the full thread sits behind a bot wall) states that the attached Revit instance “sends some analytics back”, naming which tools are called and failure information, through Revit’s regular analytics system. Per the same reply, “It cannot be turned off”, and the content of tool calls and results stays local. If a project is under an NDA that forbids even tool-name telemetry, IT needs to address that restriction before approving the add-on.

Our two prompt tests: what Claude got right, and what we could not test

On 20 October 2026 we ran two prompts programmatically from a command-line client, one run each. Claude produced a convention-correct pyRevit script with its own caveat list, and it planned an official-tool audit that stayed inside the 8 read tools our prompt listed, without inventing results. We did not execute either output inside Revit because our test machine has no Revit host; both grades are limited to static reviews.

Test 1 asked for a door-renumber script. The one-shot output had the right pyRevit anatomy: a script header, a FilteredElementCollector over the door category, phase-aware room resolution with a FromRoom fallback, and a single named Transaction with rollback. It guarded the Mark parameter against read-only states, kept a skip-and-report path for doors without rooms, and added a DRY_RUN preview flag we had not asked for.

The script came with its own warning list. Duplicate room numbers produce duplicate door marks, rooms in linked models resolve as “no room”, and it told us to test on a detached copy first. The script’s warning list did not mention that one of its pyRevit helper calls has a signature that varied between pyRevit versions; our static review caught it. Confirming that call requires running the script.

Test 2 asked for a fire-rating audit plan against the official Read tool list, which our prompt spelled out - the test checked whether Claude stayed inside that list and planned sensibly, and it did. It invented no tools, and it refused to fabricate tool results. It planned to discover how the fire-rating parameter actually exists in the model before querying, since a built-in type parameter and a firm’s shared parameter need different handling.

The plan identified two ways it could fail: wrong parameter identity, and silent scope gaps across the active view, design options and linked models. Those match the limitations in the official server’s documentation. That suggests awareness of its limits, at least on paper.

The audit prompt below is adapted from our test-2 prompt, with the model’s own classification buckets added as requirements. Paste it after connecting the Read server:

Using only the Revit MCP read tools, audit this model: which doors are missing a fire rating?
Before querying, check how the fire rating is actually stored: the built-in type parameter,
a shared instance parameter, or another name. Query the whole model, not the active view.
Classify each door: parameter absent, value empty, placeholder value ("-", "N/A", "TBD"),
or rated. Report counts per bucket, the exact parameter name you checked, and the query scope.
Then select the unrated doors in Revit so I can verify them myself. Do not modify anything.
After the summary, list the ways this audit could be wrong or incomplete in this model.

Our test archive holds both tests, their transcripts and the run log. Test 1 passed a static review; test 2 passed a documentation cross-check. In our 12-task Claude vs ChatGPT test (18 October 2026) the same model family cited a superseded code-section number while warning about that exact risk. Keep that citation failure in mind when reviewing every Revit answer.

Vibe coding: Claude writes your pyRevit scripts

Vibe coding means describing a Revit automation in plain language and letting Claude write the pyRevit or Dynamo script. It works today on any Revit version with no MCP server at all, and it is how some non-programmers on the pyRevit forum now build internal tools. Expect one or two error-fix rounds per script, and read every line that touches the model before you run it.

The pyRevit forum ran a thread on vibe-coding strategy in August 2026. Nicolas Catellier of BIM Pure wrote there on 3 August 2026: “I am personally not a coder, but I understand the Revit API (through years of using Dynamo). I think that’s enough to create basic to medium complexity tools with AI, at least for internal use.” Forum moderator Jean-Marc, in the same thread: “There is no shortcut since learning the basics is the shortcut”.

Another member keeps a bootstrap file, a skeleton script with the imports and structure he wants, and feeds it to the model with every request. He reports the output becomes far more consistent. One or two error rounds are still normal, because the Revit API changes between versions and the models blur those versions together.

One member described cross-checking one model with another. aaronrumple wrote in the same thread: “I think Claude is more solid and often use Claude to QA/QC ChatGPT code.” The thread suggests this workflow:

  1. Describe the task with inputs, outputs and boundaries. Our test-1 prompt named the category, the phase rule, the Transaction requirement and the report format.
  2. Give context: Revit version, pyRevit, and your skeleton file if you have one.
  3. Run on a detached copy of the model. Keep the live file closed while you do.
  4. Paste errors back verbatim. One or two rounds are normal.
  5. Before office rollout, have someone who can read the code review the script. If nobody in the office can, keep it personal.

The script-request prompt below is adapted from our test-1 prompt; the DRY_RUN flag Claude added unprompted is now a stated requirement:

Write a complete pyRevit script (IronPython 2.7 compatible) that renumbers all doors in the
active Revit document by the room they open into: each door gets its ToRoom number (phase-aware,
last phase), with an A/B/C suffix when a room has several doors. Requirements: __title__ and
__doc__ header; imports from pyrevit (revit, DB, script); FilteredElementCollector over OST_Doors;
FromRoom fallback when ToRoom is empty; one named Transaction for all writes; write to the Mark
parameter with a read-only guard; skip and report doors with no room; include a DRY_RUN flag
that previews the result without writing. After the script, list your assumptions and what I
must check before running it on a live model.

A generated script edits your deliverable. Review the element list it reports and keep the DRY_RUN pass in place; the responsibility for what lands on the drawings stays with the licensed professional who signs them. More tested prompts for this kind of work are in our prompt library.

What does it cost?

A working read-only setup costs nothing beyond licences you may already have. Claude’s Free plan includes Claude Desktop and desktop extensions, the official add-on is free with a Revit 2027 licence, and the open-source servers are free. Paid tiers buy usage headroom and write tools, from $19.99/month for Geopogo to 85 EUR + VAT per user per year for NonicaTab PRO.

What you can do for free:

ToolFree tierReal limitEnough for
Claude (Free plan)Claude Desktop + desktop extensions includedusage caps on a rolling five-hour windowtrying read-only queries on a small model
Official Revit MCP Serverfree add-onneeds a Revit 2027.2 licence; Technical Previewfull read-tier use on 2027.2
mcp-servers-for-revit and kinfree, open sourceyou own the security reviewtechnical offices on Revit 2020-2026
NonicaTab FREEread-only AI Connector, 36 toolsno writes, limited toolsreporting and model QA questions
IFC export + Claudeno plugin at allno live link; data is a snapshotschedules, areas, fit checks

Claude plans per Anthropic’s pricing page, read 20 October 2026: Free $0, Pro $20/month billed monthly, Max from $100/month, Team $25/seat/month billed monthly ($20 annually). Usage limits reset on a rolling five-hour window, and paid plans add weekly limits on top. Heavy MCP sessions use up the free quota quickly; for daily use Pro is the realistic floor.

On the plugin side, NonicaTab PRO is 85 EUR + VAT per user per year and Geopogo is $19.99/month, both checked 20 October 2026. Neither is required for any read-only route in this guide. Pay for them when you need write tools or vendor support.

Four ways this breaks in a real office

Four things go wrong in practice. A community server exposes your model to the office network, an AI write damages the file, a generated script fails across Revit versions, and confidential data leaves through exports or telemetry. All four have cheap mitigations: read-only first, localhost-only servers, detached copies, and an IT review of what leaves the machine.

  1. Network exposure. The most-starred community server listens on port 8080 on every network interface with no password, and it ships a tool that compiles and runs any C# inside Revit (techdrawai’s code read, 8 October 2026). On an office network that is a remote-control port for your BIM station. Prefer the official stdio add-on, which opens no network port, or a loopback-only server with tokens or pairing: bimwright/rvt-mcp started with —read-only, or the pyRevit MCP server, which shows every code payload for review in Revit (both still ship code execution by default).
  2. Model damage. The Write and Experimental tiers of the official server can modify the model, and Autodesk’s own docs say to review proposed changes and consider a backup before write operations. Claude Desktop asks approval per tool call, which helps only if you read what you approve.
  3. Version drift in generated code. The Revit API changes between releases and the models mix those releases up; the pyRevit forum reports one or two error rounds as normal even with good prompts. Our test-1 script handled the 2024+ ElementId change on its own. It still contained one call we could not verify without running it.
  4. Data leaving the machine. Model queries flow through Claude by design, so project confidentiality rules apply to what you ask and what you export. techdrawai also found that the native RVT file and the IFC exported from it carried author names and file paths nobody meant to share, and the official add-on’s tool-name telemetry cannot be turned off (Autodesk forum reply, read 20 October 2026). Check NDAs before IT signs off.

A fire-rating value read through any of these routes is whatever someone typed into the model; no route in this guide checks it against a code. The audit can miss doors in linked models, read the wrong parameter, or count a placeholder as rated. Required ratings, structural values and anything that goes on a stamped sheet stay with the licensed professional’s own verification against the current local code edition. Our building-code guide walks through that verification workflow step by step.

FAQ

Version coverage, price, model safety, Mac support, learning resources, and the Claude versus ChatGPT choice determine whether this setup works for your office. The six answers below cover these questions; the first addresses older licences, and the fourth covers Mac support. Versions, prices and Technical Preview terms were checked on 20 October 2026.

Does Claude work with Revit 2024 or 2025?

The official Autodesk server does not. It requires Revit 2027.2 during the Technical Preview. On Revit 2024, 2025 or 2026 you have three routes: a community MCP server (Revit 2020-2027 coverage varies by project), a commercial plugin such as NonicaTab (Revit 2020-2027) or Geopogo (Revit 2025-2027), or no live link at all, where Claude writes pyRevit scripts and reads IFC exports on any version.

Is the Revit MCP connection free?

Mostly yes. The official add-on is free with a Revit 2027 licence, the main community servers are free and open source, and Claude’s Free plan includes Claude Desktop with desktop extensions (Anthropic pricing, 20 October 2026). Free-plan usage caps reset on a rolling five-hour window and run out fast on MCP sessions, so daily use pushes you toward Pro at $20/month. NonicaTab FREE offers a 36-tool read-only connector.

Can Claude break my Revit model?

The official Read-Tools server cannot; its tool list has no writes. Write and Experimental tiers, and most community servers, can edit, move and delete elements, so Autodesk tells you to review each proposed change and back up first; read-only claims of other tools rest on vendor and README statements. Claude Desktop asks for approval per tool call. On a detached copy, the worst case is a discarded file.

Does this work on a Mac?

No direct route exists. Revit runs on Windows, and every live MCP connection needs Claude Desktop or Claude Code on the same Windows machine, so Mac users run a Windows VM (Autodesk documents Parallels; techdrawai and meltflexai both confirm this, October 2026). The workable Mac alternative is indirect: export IFC or schedules from a colleague’s Windows seat and let Claude analyze the export.

Is there a course for learning Claude with Revit?

Our 26 October 2026 search found no official Autodesk or Anthropic course on using Claude with Revit; the Technical Preview is too new for a formal curriculum. The closest substitutes are Autodesk’s own documentation - the Revit Public MCP Server help pages and the launch blog post, both listed in Sources below - plus independent walkthroughs such as techdrawai’s October 2026 review. The setup section of this guide is written to be followed step by step for the same purpose. Beware of unrelated offers: “Claude Certified Architect” courses teach software architecture for IT systems, not building design.

Is Claude better than ChatGPT for Revit work?

It depends on the job. For Revit API code, one pyRevit forum member’s habit is to use Claude to QA what ChatGPT writes. Our own 12-task comparison (18 October 2026) found Claude deeper on long-document judgment and ChatGPT more reliable on code-section citations, with six tasks even. For live model access, the official Autodesk server names Claude Desktop and Cursor as supported clients, which settles the choice on that route.

Sources

Every date, price and version number in this article traces to one of the sources below, read on the dates shown. Vendor pages and independent write-ups are listed side by side so you can check each claim against its origin. Prices and Technical Preview terms change; re-read the Autodesk and Anthropic pages before you quote them to a client or an IT department.

How to use Claude AI in Revit this week: on Revit 2027.2, install Read-Tools and run the fire-rating audit prompt on a detached copy; on anything older, paste the door-renumber prompt into Claude and read the script before you run it.

If this guide saved you the server-vetting afternoon, the newsletter brings one tested workflow like it every week. Subscribe.