Claude is the AI tool most non-developers in the BIM and VDC world have actually opened. The chat window is fine. It writes a passable RFI response, summarizes a spec section, and rephrases a coordination note. None of that is what is moving the needle in 2026. The thing that is moving the needle is the Model Context Protocol (MCP), and the small set of servers a coordinator or VDC manager can wire up in an afternoon.
This post is about what that actually looks like in a daily workflow on a job that is in active coordination, and where the limits are.
What MCP is, in one paragraph
MCP is a protocol that lets Claude (or any MCP-compatible client) talk to outside data sources and tools through a standard interface. Instead of pasting an RFI into the chat for the model to read, you connect Claude to the RFI tracker once, and from then on the model can look up RFIs, list open ones, fetch attachments, and cross-reference them against a sheet index without anyone copy-pasting context. The servers are typically small programs that wrap an existing API. You start one per data source, point Claude at them, and that is the integration.
The conceptual shift is the important part. The chat window is no longer the only context the model has. Claude becomes a search bar, a triage tool, and a routing layer that knows which document, model element, or person owns the answer.
The four MCP servers that earn their keep on a real job
For a coordinator running an active job, the integrations that pay back the setup time are small in number and concrete in scope.
- An RFI server. Wraps the project’s RFI tracker (Procore, Construction Cloud, Newforma, whatever). Lets the model answer “list all RFIs from the mechanical contractor that are still open and over thirty days old” without anyone running a saved view.
- A sheet-index server. Wraps the latest IFC drawing set. Lets the model answer “which sheet shows the chilled water schematic on the second floor” by mapping a description to a sheet number, then handing back the URL or local PDF path.
- A submittals / shop-drawings server. Same shape as the RFI server, against the submittal log. Lets the model answer “is the air handler submittal back yet, and what was the last comment from the mechanical engineer.”
- A local file-system server, scoped to the project folder. Lets the model answer questions against the documents the team already has on disk: spec sections, meeting minutes, coordination notes, and the latest BIM execution plan.
That set covers most of what a junior coordinator does between coordination meetings. None of those servers is hard to stand up. Most of them are existing open-source MCP implementations against the public APIs of the tools the project already pays for.
A worked example: a Tuesday morning during coordination
Consider a real prompt sequence on a hospital job mid-coordination. The team meets at 9 AM. The coordinator’s prep before the meeting used to look like this:
- Open the RFI tracker, filter by status and trade, scroll for outliers.
- Open the submittal log, find anything that was due last week and is not back.
- Open the sheet set to confirm a few sheet numbers for the meeting agenda.
- Pull the last meeting’s minutes to find action items still open.
That is twenty minutes of clicking, easily. With MCP wired up, the same prep is one prompt:
The model runs the relevant queries against each MCP server, joins the results, and returns the table. The coordinator reads it. If something is wrong, they correct it in the source system, not in the chat. The next morning’s prompt is the same prompt; the brief reflects the corrections.
Where the limits actually are
Three honest constraints that will hit any team trying this:
- The integrations are only as good as the data hygiene of the underlying systems. If the RFI tracker is full of stale items because nobody closes them, the brief will reflect that. Claude is not going to fix data hygiene. It will, however, make the consequences of bad hygiene impossible to ignore.
- Permissions and ownership matter. Pointing Claude at a project folder means Claude can read everything in that folder. That includes contract drafts, fee schedules, and the personal notes someone left in a “scratch” subdirectory. Scope the MCP server to the directories and document types that should be exposed, not the entire project root.
- Hallucination is still possible against structured data. Most MCP servers return real records, but the model can still misattribute a sheet number or paraphrase a spec section incorrectly. The rule that has worked: any output that will be sent to a contract party (RFI response, submittal redline, coordination note) gets read by a human against the source before it ships. Internal-use briefs are lower-stakes and can be trusted with a quick scan.
What this does not replace
MCP plus Claude does not replace the coordinator, the BIM lead, the project engineer, or any of the people whose job is to know what is in the model and the spec. It removes the part of their day that was spent looking up where things are. The job that is left is interpretation, judgment, and pushing on the answers, which is the work those roles were hired to do anyway.
The setup cost, plainly
Standing up the four MCP servers above on a real project, with a working AEC-stack tool integration, is about a day for someone who can read API docs, plus another day for testing against a real project’s data. The ongoing cost is monitoring the tracker tokens / API limits and updating the sheet-index server when a new IFC set drops.
That is small enough that it is worth doing on the next job that has more than ten open RFIs at any given time, which is most jobs over five million dollars.
Practical takeaway
If a VDC team has one engineering hour to invest in AI this month, do not spend it on a Claude prompt-engineering session. Spend it standing up an MCP server against the RFI tracker. Everything else builds on top of that one integration: the sheet-index server gets used because the RFI server made cross-references easy, and the submittal server gets used because the brief grew to want it. The chat window is the front door. MCP is what makes the rooms behind it useful.