Iam a mobile developer i dont have a pc Or laptop,but iam a passionate developer and loves to code and has good analytical skills,means I develop this whole webmcp through my mobile with arena agent and fix error, vulnerability with qodo code review tool, my webmcp server has almost no error, bugs and vulnerability which typically ai agent do and my webmcp are the one of the best webmcp that you are going to see in this hacakathon It's has physics 40 different tools And ai agent freedom. Means I put everything an webmcp of 3 d model can actually have and done the best impact it could possibly do.
Inspiration
Most ai are like can control only the simple tools and has acess of few simple features but we don't want this simple. That's why we build orbit Orbit uses edit_geometry to plan excited 3 d model it uses all the tools and features provided by the mcp server not only the basic feature, it uses all the features and can implement any thing that he wanted to.
What it does
Orbit is a browser based 3 d studio or an mcp server that uses ai+human to build the 3d model, and has smart technology and workflow Our core workflow is: Observe → Plan → Propose → Approve → Act → Verify
How we built it
Building orbit is not an easy task it need preparation,first we has to build the working prototype then we build the core working system and then we designed tools that ca be exposed too ai agent then we refine the architecture,dataflow,clearout all the vulnerability,error and bugs. Orbit is built with plain JavaScript ES modules, with no framework or build step. A single shared scene store powers the viewport, inspector, AI planner, Webmcp tools, and persistence. This keeps the human and AI working on exactly the same scene state. The natural-language router is a pure, deterministic function that converts user requests into tool calls. This makes the agent predictable and easy to test. Our current router evaluation is 38/38 passing.
Challenges we ran into
There were many challenges in the orbit like the early stl exports could appear correct but fail in slicers because some geometry had zero thickness. We fixed this by deriving the export from the actual scene geometry.One challenge was balancing live AI updates with reliable undo. We solved this by grouping an entire AI run into a single undoable operation.
Accomplishments that we're proud of
Iam proud of many things that I have done that Orbit turned AI-controlled 3D editing into an actual human + AI collaboration system, rather than a simple AI demo. We're especially proud of: 32 WebMCP tools 38/38 routing evaluations passing Approval-gated actions Real solid STL generation Geometry validation Undoable AI operations
What we learned
We learn that ai doesn't need a better autonomy, it needs that better resources, interface and the clear boundary
What's next for Orbit
We want to make orbit more realibale with many tools and more efficient in building complex task that I think we have done when you are seeing our submission we are adding more tools not only simple one we want that the ai agent can modify object like they want every minor to minor detail.
Log in or sign up for Devpost to join the conversation.