Revit MCP: How AI Agents Read and Write Your Model

You have an AI assistant that can write a schedule formula, explain a Revit API error, and draft your BIM execution plan — but the moment you ask it something about your model, it goes blind. "How many unplaced rooms are on Level 3?" It can't know. So you do the dance: open a schedule, export to Excel, paste the table into the chat, and hope you copied the right columns. The model holds the answer; the assistant just can't reach it.

Model Context Protocol — MCP — is the standard that closes that gap, and as of mid-2026 Revit has an official one. This article explains what Revit MCP actually is, how to connect an AI to your model today, what it can and cannot do yet, and where the real frontier — letting an agent act on the model, not just read it — is headed. Everything in the read section works right now.

What MCP is, without the hype

MCP is an open standard for connecting AI assistants to external tools and data. Think of it as a USB port for AI: instead of every assistant building a custom integration with every application, MCP gives them one common plug. A program called an MCP server exposes a set of tools — structured functions like "find elements" or "read parameter" — and any MCP-capable client, Claude for instance, can call them.

For Revit, that means an assistant can ask your model direct questions and get structured answers back: no copy-paste, no spreadsheet export, no describing your model in a paragraph. The assistant doesn't "look at" your screen. It calls a defined function, the server runs that function against the live document, and returns real data. Hold onto that distinction, because it's the whole game: an MCP tool is deterministic code, not a guess.

Revit's official MCP server (Tech Preview)

In June 2026 Autodesk released the Revit Public MCP Server as a Tech Preview. It's real, it's usable today, and it's the sanctioned place to start. Out of the box it lets an AI assistant:

• Find and count elements — "how many doors on Level 2", "list every in-place family", "which views aren't on a sheet".
• Read parameters — pull values across a selection or category without first building a schedule.
• Make bulk parameter edits — set a shared parameter across many elements from one instruction.
• Capture view snapshots — hand the assistant a visual of a view to reason over.

Setup, briefly: install the server for your Revit version, point an MCP client (Claude Desktop, or another MCP-capable app) at it, and the tools appear inside your assistant. From there you ask in plain language and the assistant picks the tools. Autodesk is explicit that this is a preview and will evolve — but the read layer is genuinely useful now.

Two honest caveats. First, the preview targets recent Revit (the 2027 line) — confirm version support before you plan around it. Second, there are community, open-source MCP servers for Revit on GitHub too, some with dozens of tools, and a few are excellent — but they're unofficial. You're trusting a third party's code with access to your model, so vet them the way you'd vet any add-in before you grant write access.

What the read layer solves — and where it stops

Give an assistant the read tools and a surprising amount of coordination work gets faster. "Which rooms lost their bounding on Level 4?" "List every CAD import and its file size." "How many warnings, grouped by type?" These are questions you'd normally answer by building a schedule or running a report. Now you ask them in a sentence and get a table back, in context, in the middle of the conversation you were already having.

But notice what all of those share: they read. The official preview reads, counts, and edits some parameters — creating and managing elements is deliberately held back. Autodesk has said full create-and-modify capability will arrive through a separate, dedicated write server, keeping read and write concerns cleanly apart so you control exactly what an agent is allowed to change.

That separation is the right call, and it's also where the interesting problem lives. An assistant that can tell you there are twelve unplaced rooms but can't help you resolve them is a very smart dashboard. The moment it can act — safely, predictably, with your approval — it stops being a dashboard and becomes an assistant.

The write frontier — and how B45 approaches it

The write side is the gap B45 Labs | Connect is built around. Connect is our MCP-based interface to the B45 command suite, and its premise is deliberately unglamorous: don't have an AI improvise changes to your model — have it call the same deterministic commands you already trust from the ribbon.

Concretely, B45 tools like Create Sheets, Check Model Health, and the QA/QC checks are already engines that do one defined job the same way every time. Connect exposes those engines as MCP tools. So when you say "create the sheet set from this index," the assistant doesn't invent a sequence of API calls — it calls the same Create Sheets engine you'd click, with the same title-block handling and numbering logic. Ask it to audit the model and it runs the same deterministic checks you'd run by hand. The AI decides which tool and when; the tool itself is boring, tested code. That is the promise made literal: ask Claude to audit your model, and it calls the checks you already rely on — no magic, no guessing.

Two decisions matter for trust. Writes are gated — an action that changes the model surfaces a confirmation step before it commits, so the agent proposes and you approve. You are never one loose sentence away from fifty new sheets. And every tool relays the command's own report back into the chat, so you see exactly what happened, in the same format the ribbon command would have shown you.

Honest status: Connect is in active development, not a public download yet. It ships on the B45 roadmap alongside the 2027 line, and it will follow the same free-first approach as the rest of the suite. The read layer above you can explore today with Autodesk's preview; the write side is what we're building toward — and this article is the flag in the ground while the category is still forming.

[SCREENSHOT: B45 Connect inside a chat client — a natural-language request ("create sheets from this index"), the Create Sheets tool call, and its result report, dark theme]

Frequently asked questions

What is MCP in Revit? MCP (Model Context Protocol) is an open standard that lets an AI assistant connect to Revit through defined tools — functions that read or change the model. Autodesk's Revit Public MCP Server (Tech Preview, 2026) is the official implementation: it lets assistants like Claude query your model directly instead of you exporting data by hand.

Is Revit MCP safe to use on a real project? The official Autodesk server is read-focused and clearly scoped, which limits the risk. The care is all on the write side: only let an agent change your model through tools that are deterministic and gated behind an approval step, and treat any third-party or community MCP server like any other add-in — read the code before you grant it write access.

Can an AI actually create elements in Revit? Not yet through the official preview, which reads and edits some parameters but holds element creation back for a future dedicated write server. Purpose-built approaches like B45 Connect target exactly that frontier by exposing tested command engines — for example Create Sheets — as gated tools, rather than letting the AI improvise API calls.

Do I need Revit 2027 for MCP? The official preview targets recent Revit, so confirm version support before you plan around it. MCP as a protocol is version-agnostic; what decides compatibility is which server supports your Revit version.

Start where it's real today

The read layer is here and worth an afternoon: set up Autodesk's Revit Public MCP Server, point Claude at it, and ask your model the questions you used to export spreadsheets for. When the write side matures — ours and Autodesk's — you'll already know how to work this way.

B45 Labs' Revit tools are free until 2027 — 50+ commands, no credit card. Download the suite and start with Check Model Health; then read the companion piece on running a full model audit for the checklist an agent will one day run for you.