Revit Agents is a governance and orchestration layer for AI agents connected to Autodesk Revit through MCP. It gives BIM teams control over what Revit AI agents can access, execute, approve and audit — while Revit, your models and the local LLM stay on the workstation behind a secure connector.
A BIM Execution Plan gives model delivery a shared, agreed method. Revit Agents applies that same mindset to Model Context Protocol automation — it does not replace your contractual BEP.
Control plane
The browser never talks to Revit. Approvals travel down a paired, authenticated path; execution happens locally.
Capability is detected per workstation — the console reports what is genuinely reachable rather than assuming a live connection.
A Revit agent is an AI agent that can read from and act on an Autodesk Revit model through a defined set of tools. Those tools are published by an MCP server running beside Revit on the workstation, so the agent never invents capabilities — it can only call operations that have been exposed to it.
Revit MCP is the capability layer: an MCP server for Revit that exposes status, query and write tools to a language model. Revit Agents is the governance layer on top — who may run which agent, with which tools, against which project.
Revit Agents vs Revit MCPBIM automation agents fail in specific ways: over-broad tool access, unintended model edits, ambiguous classifications accepted silently, destructive actions, stale context and thin audit trails. Governance addresses each before an agent is allowed to modify a live model.
Read the governance guideTool access, model context, prompt versions, approval gates, preview versus applied execution, and the audit trail left behind by every run — scoped by company, project and agent.
Agent tools explainedLocal MCP proves the automation works. Governance is what makes it safe to operate across a practice — the same step BIM took when individual modelling habits became an agreed execution plan.
Excellent for proving the idea.
Ready to operate at practice scale.
Policy is set once by the practice, tuned per project, and executed locally.
One approved method, set once by the practice.
Configured to the delivery standards of each job.
Execution never leaves the machine running Revit.
Every stage is explicit. Nothing is written to Revit until an authorised run is approved, and the result comes back to be verified and recorded.
Company standards, permitted tools and safety rules.
A prompt version moves from draft to active.
Scoped configuration for that job's standards.
Read-only inspection of the proposed change.
The Revit write happens only after approval.
A verified result, recorded in run history.
Analysis and preview are read-only. The Revit write only happens on an authorised run, and returns a verified result to the audit record.
From the console to the Revit model and back — every hop is authenticated. The browser never reaches your workstation directly: Cloudflare Access verifies the service identity, and the local connector checks a separate pairing token before anything runs. Scroll to watch the path light up.
Practical controls that make automation reviewable. Some are in place today and others are actively being built out during early access — we will be explicit about which is which in the console.
Who may configure agents, who may approve, who may run.
Agents are granted only the tools their task needs.
Editing creates a new version; an active prompt is never silently overwritten.
Analysis output is clearly distinguished from changes applied to Revit.
Runs are routed to a specific paired workstation through the secure connector.
Settings and approvals do not leak between projects.
Designed for a durable, append-only record of what ran and who authorised it.
Automation that a practice can stand behind — and that security teams can reason about.
Practical background for architects, BIM managers, computational designers and Revit API developers evaluating safe Revit automation.
Anatomy of an agent, where it runs, and what it can and cannot do.
How an MCP server for Revit exposes tools to a local LLM.
Capability layer and governance layer, and how they combine.
Approvals, permissions, prompt versions and audit trails.
The local workstation boundary, connectors and least privilege.
Drylining classification, standards audits, schedules and sheet QA.
Status, read, preview and write tools — and how to grant them.
Why workstation LLMs keep model data inside the practice.
Early access is opening to a limited number of practices. Join the waitlist and we will be in touch as places become available.
Already invited? Sign in
What is Revit AI? How local LLMs keep your models safe