Autodesk
MCP Servers
The Situation
The team: 10 working groups, every major product at Autodesk
My role: Experience Architect
MCP Servers were changing what AI could do - threatening the relevancy of the assistant Autodesk was building. I switched out of management to design the ecosystem for Autodesk's company-wide MCP Server program in 6 months.
Timeline: 6 months
A pivotal moment
Spring 2025 - MCP servers were taking the industry by storm. Autodesk could either build on their robust ecosystem of APIs, or have it completely disrupted by overuse. Autodesk's assistant still couldn't take action on its own, only answer questions, like a fancy help guide.
A good MCP server helps AI agents connect user requests to available API functionality. They're built on a strong use case, and a deep understanding of how customers talk about their problems..
After managing for 9 years, I was ready to move back into an individual contributor role, and I jumped at the opportunity to partner with the company's best architects and strategists to roll out this program and get it to market - in 6 months!
End-to-End design
The MCP Server ecosystem would include internal and external developer experiences. We needed to provide for product discovery via in-app experiences, listings in our e-store, app store and other channels, such as Claude marketplace. It would fundamentally change certain parts of our agent experience. And it would require a suite of administrative controls and the ability to customize to individual company's requirements.
I successfully coordinated among teams so we could announce and do live demos onstage at our large annual customer conference in mid-September. This included partnering intensively with Autodesk's Assistant team, who builds the primary agent interface, and working closely with our design system team to audit patterns and craft solutions for the numerous changes required to support MCP servers.
Future of AI + CAD
MCP servers and generative AI will enable a whole new way of working with CAD. Autodesk's interfaces are notoriously difficult to learn. My VP challenged me to prototype what the future would look like 3-5 years out. I worked with experts in each industry team to identify meaningful workflows to test. During this time I was also appointed to advise our research team, who were building for 10 years out on a project called JARVIS. I coordinated with them to ensure my research would feed valuable early customer data into their project.
I built and tested a prototype showing 5 different experiences:
-
Reviewing of massive amounts of agent actions (with Michelangelo Caparo)
-
A 3D canvas with spatial AI context building
-
Minimal UI that centers the canvas experience
-
Collaborative, multi-agent experiences
A sample daily readout to my all-remote team while prototyping during an offsite with the JARVIS research team. This represents 4 hours of prototyping using Lovable and exchanging ideas back and forth with a developer in the UK, who built a 10,000 node map in the back end to test the performance and feasibility of our architecture.
This video gives a good idea of my daily output. We spent about 4 hours specifying our idea and researching the problem, and 4 hours prototyping.
Outcomes
We recognized that working with AI is much like working with human collaborators. At some point soon, we may work with multiple agents as our team.
The reaction to this prototype was strong - people were ready to bring agents into their workflows, and they particularly wanted ways to intuitively build their project context. The canvas could become the way to do so. Another team is now building that product.
We continued to build on this idea with the JARVIS research team, looking at building massive canvases with world models.
