MCP for Revit: connecting AI assistants to model data

What the Model Context Protocol is, what a Revit MCP server would actually have to do, which workflows it would be useful for, and what it must never be allowed to do. BIMRACE has not built one — this page explains the architecture rather than selling it.

BIMRACE status
Roadmap — not built
What is live
Automation routines
What is in development
Model data extraction
Write access
Not proposed

Read this first. BIMRACE does not have a working Model Context Protocol server for Revit. We have not shipped one, we do not operate one, and nothing on this page is a product you can buy or licence. Roadmap is the honest badge for the protocol layer and it is the badge it carries in our navigation.

What is real today is underneath it: automation and checking routines that have been Live on client appointments for years, and model data extraction and reasoning that is In development and used on our own delivery. This page describes how those would connect to an AI assistant through MCP, because it is a question worth answering properly in a market that is currently answering it loudly and badly.

What it is

A declared interface, not a smarter chatbot

The Model Context Protocol is an open standard for connecting AI assistants to external tools and data. A server exposes a set of typed tools and resources; a client discovers what is available and calls it. The protocol itself is unremarkable — which is the point.

What matters for engineering is the consequence of that structure: the complete set of things an assistant can do against your model is a finite, enumerable list that somebody wrote and somebody else can review. Compare that with the alternative currently being sold, which is an assistant with unspecified access and a reassuring paragraph about safety.

An engineering practice should be able to read the tool list, disagree with an entry, and have it removed. That is a governable interface. An integration you cannot enumerate is not.

ARCHITECTURE / WHERE MCP WOULD SIT Roadmap
  1. Engineer

    Asks a question

    In their own words, about a model they are accountable for.

  2. Assistant

    Selects a declared tool

    From an enumerable list, not from unspecified access.

  3. MCP layer

    Marshals the call

    Into Revit's API context, inside a read-only boundary.

  4. Model

    Returns structured data

    Elements, parameters, systems, relationships — with element IDs.

  5. Engineer

    Reviews and decides

    Nothing is committed to the model by anything above this line.

READ BOUNDARY — NO WRITE PATH PROPOSED ARCHITECTURE DESCRIPTION, NOT A RELEASED PRODUCT

The hard part

Why a Revit MCP server is not a wrapper around the API

Most published enthusiasm about AI and Revit skips this section. It is the section that determines whether any of it works on a real project.

Constraint 01The API is not a service
Revit API calls must execute on Revit's main thread inside a valid API context. A server receiving a request cannot simply call the API from its request handler; every call has to be marshalled into Revit's execution context and the result handed back. This is solvable with external events and an idling handler, and it is why a Revit MCP server is an application hosted inside Revit rather than a process beside it.
Constraint 02Transactions and document state
Any model modification requires a transaction, and a transaction opened by a background process against a document a person is actively editing is a way to lose someone's afternoon. Read operations have their own version of this problem: a query answered against a document mid-edit may not describe anything that will exist a minute later.
Constraint 03No headless interactive mode
Revit is not designed to run headless against interactive documents. That constrains the deployment shape considerably: the server lives where Revit lives, on a workstation or a controlled session, not as a scalable cloud service answering questions about a model that is not open.
Constraint 04Model size versus context
A federated MEP model contains far more data than can be handed to a language model. Any useful implementation is therefore mostly a query and summarisation problem — resolving a question to a narrow, structured extract — and only incidentally an AI problem. Tools that return whole model dumps are demonstrations, not workflows.
Constraint 05Non-determinism against engineering
A language model will occasionally return a confident answer that is wrong. In a conversational product that is an annoyance. Against a model that a duct size will be taken from, it is a defect with a professional liability attached, and it is the reason the read boundary and the engineer review step are architectural rather than optional.

Use cases

What it would genuinely be useful for

Ordered by how close each one is to being real. The badge on each is the status of the underlying capability at BIMRACE, not of the MCP layer — that remains Roadmap throughout.

01

Model interrogation In development

Element counts by system, parameter completeness by family, where a given type is used, which elements are missing a required value. Today this is a script somebody has to write for each question. The value of a declared tool interface is that the question no longer needs a script.

02

QA interrogation Live

Asking why a QA run produced a finding, which rule fired, and what else shares the same root cause. The rule sets and the findings are live today; what is missing is the conversational route into them.

03

Coordination questions Live

Which findings belong to which discipline, what changed since the last federation, which issues recurred. The coordination data exists; querying it in plain language does not.

04

Quantity queries Live

Quantities by system, by level, by classification, with the measurement rule and the element list behind each line. Automated extraction is live; asking for a cut of it is not.

05

Documentation drafting In development

Drafting the repetitive parts of a report, a schedule note or an issue description from model state, for an engineer to correct rather than to write from nothing.

06

Onboarding a model Roadmap

Answering "how is this model organised" for someone who has just inherited it — worksets, naming, parameter schema, system structure, where the conventions have been broken. Not built, and arguably the most valuable of the six.

The read boundary is the design. Every use case above is a read. That is not a phase-one limitation with write access scheduled for phase two — it is the boundary we would build to and stay behind.

A routine that silently edits a live model removes the one thing that makes an engineering deliverable worth anything: a person who is answerable for it. Automation here drafts, proposes and reports. It does not commit.

Where an engineer signs. The review step is not a checkbox at the end. It is the point at which a named person takes responsibility for an output, and it is why the architecture keeps every AI-produced artefact in a reviewable state rather than in the model.

What it must never do

The list is short, and none of it is negotiable

  • Write to a live model without a named engineer accepting the change first.
  • Close a coordination issue, change a status code or issue a deliverable.
  • Produce an engineering value — a size, a load, a rating — presented as authoritative rather than as a draft for checking.
  • Send model data outside the environment the client agreed to, or retain it afterwards.
  • Operate against a document somebody else is actively editing, without their knowledge.
  • Present a probabilistic answer with the confidence of a deterministic one.

Every item on that list is currently being done somewhere in this sector, usually in a demonstration video, usually with the reviewing engineer edited out.

Security

The questions to ask any vendor proposing this, including us

These are not rhetorical. They are the questions that decide whether an AI integration is adoptable by a practice with NDAs and professional indemnity, and most current offerings answer them badly or not at all.

Question 01What leaves the machine?
Which model data is transmitted, to which service, in which jurisdiction. "It is secure" is not an answer to this question; an architecture diagram is.
Question 02What is retained, and by whom?
Whether prompts, extracts or model data are stored, for how long, and whether they are used for training. A client NDA that predates the AI strategy does not stop applying because the tool is new.
Question 03Can the tool list be enumerated and restricted?
If the set of operations an assistant can perform cannot be listed and reduced, the integration cannot be governed, whatever the marketing says.
Question 04Is there an audit trail?
Which tool was called, against which document, by whom, and what it returned. Without this, an incident cannot be investigated and a finding cannot be defended.
Question 05What happens when it is wrong?
Not whether it can be wrong — it can. Whether the workflow has a review point designed into it that catches a wrong answer before it reaches a deliverable, and who is named at that point.
Question 06Does it degrade safely?
If the service is unavailable or the answer is refused, does the engineer's workflow continue, or has a dependency been introduced into the critical path of issuing a deliverable.

Where we actually are

The honest position, stated as three lines

LayerAutomation and checking routinesLive
Running on client appointments now — model QA rule sets, clash classification, quantity extraction, documentation generation, parameter management. This predates the platform framing; it is how the work already gets done.
LayerModel data extraction and reasoningIn development
Elements, parameters, systems and relationships read as structured data, with pattern and anomaly detection over it. In development, used internally on our own delivery, not released and not licensable.
LayerMCP server for RevitRoadmap
Not built. Published here because the architecture is coherent and the constraints are worth writing down, not because it exists. No date is committed, and if that changes this page changes with it.

If you find a claim on this page that we cannot support at enquiry stage, that is a defect and we want to hear about it. The four-state legend behind every badge is published at capability status.

FAQ

MCP and Revit — the questions worth asking

Including the one about whether this is real, which is answered first.

Does BIMRACE have a working MCP server for Revit?

No. This is marked roadmap on this page and in the platform navigation, and nothing on this page should be read as a released integration. What exists today is model data extraction and reasoning, which is in development and used on our own delivery, and the automation and checking routines that have been live on client appointments for years. The protocol layer between them and an AI assistant is not built.

What is the Model Context Protocol?

An open protocol for connecting AI assistants to external tools and data. A server exposes a set of typed tools and resources; a client — an AI assistant — discovers them and calls them. The value is that the interface is declared and enumerable rather than improvised, which means the set of things an assistant is able to do against a system is a finite list somebody wrote down and can review.

Why is Revit a difficult target for this?

Because the Revit API is not a service. Calls have to run on Revit's main thread inside a valid API context, model changes require transactions, and Revit is not designed to run headless for interactive documents. An MCP server cannot simply call the API from a request handler — it has to marshal every call into Revit's execution context and hand the result back. That is an engineering problem with known solutions, but it is the reason a Revit MCP integration is more than a wrapper.

Would it be able to change our models?

Not in anything we would build. The read boundary is the design: query, extract, report and draft. Write operations against a live model would remain behind the same constraint as every other automation on this site — a named engineer accepts, amends or rejects, and nothing commits without that. This is not a technical limitation we expect to remove later. It is the thing that makes an engineering deliverable worth anything.

Does model data leave our environment?

That is the question to ask any vendor proposing this, and the answer determines whether it is adoptable at all. A model contains commercially sensitive design information and frequently sits under an NDA that predates anyone's AI strategy. Any architecture we would propose has to be explicit about what leaves the machine, what is retained, and by whom — before it is interesting rather than after.

Can we talk to you about this anyway?

Yes, and it is a more useful conversation than a demonstration would be. We can show the extraction and checking work that is real today, describe honestly where the protocol layer would sit, and tell you which parts of what you have read elsewhere are further away than they are being presented as.

Engineering enquiry

Evaluating AI for BIM? Ask for the working session, not the demo.

We will show the extraction and checking work that is real today, describe honestly where the protocol layer would sit, and tell you which of the claims you have read elsewhere are further away than they are being presented as.

Email
info@bimrace.com
Telephone
+91 75079 58364
Response
Within two working days
Reaches
Somnath Baste, Founder