McNeel released Rhino 9 in late October, and its AI feature group arrived with a design decision worth separating from the feature list: the company has built the bridge between an AI assistant and a live Rhino document, published it under a permissive licence, and declined to supply the assistant.
McNeel’s own page for the release lists RhinoAI as one of thirteen feature areas, alongside freeform modelling, 2D drafting, Grasshopper 2 and the development platform. Its description is short, and most of it is about who is in control. RhinoAI “connects an AI assistant to your open Rhino and Grasshopper document.” You can use “Claude, Codex, another assistant you already use, or a model running on your own machine.” And: “It can only use the tools you allow.”
AEC Magazine’s report from Rhino World 2026, published 12 November, is where the rest of the picture comes from. The magazine puts the event at 21–23 October near Barcelona with about 330 attendees, dates the Rhino 9 release to 28 October for Windows and Mac, and counts more than 100 new features grouped into thirteen “stories”. It also reports the thing a release page would not: RhinoAI was built by Callum Sykes, using Claude Code.
The Shape of the Integration
The code is public. McNeel’s GitHub repository describes itself
as “AI features for Rhino” and the README as “An AI Integration for Rhino3d”. It is MIT-licensed,
tagged agent, ai and mcp, and the branch on display is rhino-9.x with 913 commits. At the time
of writing it carries 407 stars and 67 forks, with 20 open issues and 13 open pull requests.
The README’s quickest install route is through Claude Desktop: a connector.mcpb connector file,
version 0.1.6, which installs the Rhino plug-in itself. The listed clients are Claude Desktop, Claude
Code, GitHub Copilot, OpenAI Codex, Gemini CLI and a local model. AEC Magazine adds that the in-Rhino
panel uses agents already installed on the user’s machine, that outside Rhino a user can bring their
own model and agent, and that RhinoAI is a plug-in rather than part of Rhino’s core — users choose
whether to install it and carry the model costs themselves. The two framings sit together rather than
against each other: McNeel lists it as a Rhino 9 feature area, and it is distributed as something you
opt into.
That last point is the commercial one. This publication has now covered a long run of AEC vendors shipping MCP servers, and the question that separates them is who pays for inference and who picks the model. Qonic metered its in-model agents by the token, separately from a platform priced by portfolio size or per user. Graphisoft’s product page listed an Archicad MCP as beta alongside a new Archicad Max plan for “AI & data driven practices”, with no price attached to any of its three tiers. Chaos published MCP servers for V-Ray, Cosmos and Anima reachable through a single V-Ray Collection licence. McNeel’s answer commits the company to least of all of them: the bridge is free and open, the model is the user’s problem, and nothing about the arrangement requires McNeel to run a token meter or make a frontier-model procurement decision on behalf of its customers.
It is not the first MIT-licensed MCP package in this market — a Sydney builder’s software published its own under MIT in September. What is distinctive here is the combination: a desktop CAD vendor shipping the integration as part of its flagship release, under that licence, while naming a model running on the user’s own machine as a supported option. That last part is the configuration an architecture practice working under a client confidentiality clause will be asked about.
”It Can Only Use the Tools You Allow”
That sentence is a permission model stated in a sales page, and it is the right place for it. The write-access problem is the live issue in this category: an assistant that reads a model is a search box, and an assistant that changes one is a colleague with no training record. This publication has tracked the same tension in AEC firms governing vibe-coded internal tools and Revit MCP write access and in ProjectTeam scoping which records its MCP server may write to.
What McNeel’s page does not do is enumerate the tool set, say whether permissions are per-tool or
per-session, or state whether the default is allow or deny. The repository is more forthcoming than the
sales page: its PRIVACY.md names run_command, run_python and run_csharp as tools that execute
commands or code inside the Rhino process, and open_doc and save_doc as tools that read and write
files, and tells users to “treat the connector with the same caution they would any tool that runs
scripts on their machine.” That is the sentence a practice evaluating this needs, and it is in the
repository rather than on the release page. Read as analysis rather than reporting: a permission
claim is only as good as its granularity, and “the tools you allow” is a promise whose value depends
entirely on how small a thing a tool is.
McNeel Used the Same Tools on Itself, and Published Numbers
The part of AEC Magazine’s report that will travel furthest is not about RhinoAI at all. The magazine reports that McNeel now uses AI internally for development and testing, and gives figures for what that did to output. Developer Steve Baer told his management team he expected a 20% improvement in productivity; the magazine’s line is that “what they got was closer to three times.” The figures behind it belong to Baer and Andy Le Bihan, McNeel’s head of visualisation development, jointly: between October 2022 and June 2026 the two of them averaged 3.3 merged pull requests per working day, and from August to mid-October the figure was 9.5. Their monthly count of fixed Rhino 8 issues rose from an average of 15 in the first half of the year to 28 over the summer. RhinoCommon gained more than 150 public types and 2,200 public declarations, and an overnight automated process reads crash dumps, attempts reproduction and fixes, and tests the results.
Those are McNeel’s figures, reported by AEC Magazine, and this publication has not independently verified them. Two things are worth saying about them anyway. The first is that they are unusually specific for this genre — a merged-PR rate with a named baseline period is a harder number to publish than a percentage, and a company that publishes one has committed to a measurable. The second is that merged pull requests measure throughput, not value; a codebase can absorb three times the commits and ship the same software. The crash-dump loop is the more interesting claim, because triage is exactly the kind of bounded, repetitive, verifiable work where an automated attempt costs little when it fails.
Alongside it, the magazine reports that Brian James, in McNeel’s training and technical support team, published four free MIT-licensed plug-ins through Rhino’s Package Manager, Woolly Rhino among them, and built them with Claude Code connected through an MCP server to a live copy of Rhino — and that his slide opened with “I don’t write or read code.”
Rhino.Compute on Linux, Priced
Another item in the report comes with a price attached. Rhino.Compute is what McNeel describes as a way to “Access the Rhino and Grasshopper SDKs through a stateless REST API”, and it now runs on Windows and Linux servers. AEC Magazine reports the Linux build installs through a standard package manager, is also available as an AWS Marketplace image, and is billed at $0.10 per core-hour — a 2-core server running eight hours a day for a month costing $48. McNeel’s own core-hour billing guide quotes the same rate and the same $48 example. The magazine reports a claim that this runs roughly half to two-thirds the cost of equivalent Windows servers on Azure and AWS, and describes the Linux build as a work in progress not yet intended for production, with third-party plug-in loading still being tested.
For AEC readers the relevance is indirect but real. Grasshopper definitions running server-side are how computational design teams expose geometry logic to a web app or an internal tool, and this publication has reported on AEC firms building such tools in-house and then having to govern them. A cheaper compute tier and an agent that can drive the desktop application are the two halves of the same shift.
Bob McNeel’s line at the event, per AEC Magazine, was “Nothing’s changed, everything has changed.” The AI posture in Rhino 9 is consistent with that reading: adopt the tooling aggressively in-house, ship the bridge and give it away, and leave the model, the bill and the permissions with the customer.
