I Controlled Revit From My Phone by the Pool — Using Codex
From a pool in Bangkok I gave one spoken instruction to Codex on my phone, and the machine upstairs built two full Revit drawing sets — twelve issued sheets across two rooms — stamped the issuance, and updated the drawing register itself. Voice was the interface. The bridge did the work. Revit stayed the source of truth. Not fully headless yet — but the desk stopped being the bottleneck.
Watch the full walkthrough: https://youtu.be/JOinJvGBM-E
TL;DR: From a pool in Bangkok I gave one spoken instruction to Codex on my phone. The machine upstairs built two full Revit drawing sets — twelve issued sheets across two rooms — placed the views to the practice's templates, stamped the issuance, and updated the drawing register itself. Voice was the interface. The bridge did the work. Revit stayed the source of truth, and nothing committed until I approved it. It is not fully headless yet. But the desk stopped being the bottleneck.
I just asked for two full drawing sets. I was not at my desk. I was at the pool.
The computer doing the work was upstairs. I was not.
For years people have wanted this. Trigger Revit from a phone, kick off real work, and walk away from the machine. There are companies that sell it — Revit automation on rented virtual machines, priced for the enterprise.
This was a phone, a pool, and a computer I already own.
What actually happened
Here is the whole thing, honestly.
Revit runs on the machine upstairs. The bridge I have been building connects it to an AI client through MCP. Codex can now reach that machine, so it can drive the bridge.
Voice is just the interface. Underneath, it is the same bounded tools reading and writing the model. The model stays the canonical source of truth. Nothing about the integration layer changed — only where I was standing when I triggered it.
Revit is still open up there. So no — this is not fully headless. But I did not have to be in the room, and that is new.
What got built
Two rooms. Six drawing types each: general arrangement, finish, furniture fit-out, electrical, reflected ceiling plan, and elevations. Twelve sheets in total.
The views were placed to the practice's existing templates — not invented, placed. The same template naming the practice already uses every day, driven through the same sheet automation tool the practice already relies on. The AI did not design a new way to document a room. It followed the one already defined.
The proof: issuance and the register
This is the part that usually gets done by hand at the end of a long day.
Each sheet received its issue date and description. Then the drawing register updated itself to match — rows appearing, sheet numbers and issue status filling in.
That is the difference between a picture of a drawing set and an actual one. A render can show twelve nice-looking sheets. It cannot stamp real issuance and reconcile a register. The register is the proof that this is a documentation system, not an image.
Governance did not move
This is writing to the model, so the rules still apply.
The bridge prepared the set, stamped the issuance, staged the register — and stopped. Nothing committed until I approved it from the phone. Write tools stay disabled by default. The allowlist is explicit.
The allowlist holds whether I am at the desk or at the pool. A relaxed chair does not earn the AI more permission.
An assistant that can quietly issue drawings on its own is not convenience. It is a liability.
Where this goes
This is the part I am actually chasing.
If a phone can trigger two rooms, it can trigger a level. Give the instruction the night before. Walk in to an issued set and a register that already matches.
It is not all built yet. The next phase is project-type playbooks — workplace interiors first, then the other sectors the practice works on. Each project type carries its own conventions for sheet sequences, schedules, legends, numbering, and details. The playbook makes those conventions legible to the AI, so the instruction can move up a level — from prepare these six sheets to build the documentation set for this room, and eventually for this level.
But the hard parts — the connection, the issuance, the register — those are working today.
Reflection
Headless Revit isn't here. This is an honest step toward it.
What changed is small to describe and large in practice: the desk stopped being the bottleneck. So did the office. The work still ran on a real machine, against a real model, under the same standards and the same allowlist as always. I just was not the thing standing between the instruction and the output.
That is what the integration layer is for. Not a smarter render. A connected system that does real, issued, registered work — and asks permission before it commits.
More from the lab soon.
Key Takeaways
- I triggered two full Revit drawing sets — twelve issued sheets across two rooms — from my phone at the pool, through a bounded bridge. Codex was the interface. Revit stayed the canonical source of truth.
- Every sheet carried real issuance, and the drawing register updated itself to match. That is the proof it is a documentation system and not a render. Nothing committed until I approved it from the phone.
- Governance did not move with me. Writes are allowlisted, staged, and confirmed. The allowlist holds at the desk or at the pool.
- This is not fully headless Revit — the machine still runs Revit. What changed is that the desk stopped being the bottleneck. Enterprise platforms rent virtual machines to do this; this ran on a computer I already own.
- Next phase: project-type playbooks, so the instruction can move from prepare these sheets to build the documentation set for this level.
Headless Revit isn't here yet. But the desk stopped being the bottleneck — and so did the office.