I came into this project quite randomly and naively. I had just finished my degree and felt an urge to build something useful in robotics. The idea for Mecha Labs came from my frustration with repeating the same setup work across robotics projects. Software had gone through a huge shift in developer tooling during the 2000s and 2010s, and I felt robotics needed something similar.
I started with a web-based renderer built on WebGL and Three.js. It could open URDF, MJCF and OpenUSD robot models directly in the browser, with no local setup. I open-sourced it and shared a demo. It reached about 200 GitHub stars, which gave me the confidence to expand it into a platform for sharing open-source robotics projects.
The first version of the browser-based renderer.
Along the way, a Discord community of about 300 builders formed around the project.
Under the hood, the renderer used Three.js with
urdf-loaderfor URDF, MuJoCo compiled to WebAssembly for MJCF, and an OpenUSD web viewer. I built the first version in a couple of days, so the pieces didn't fit together super cleanly. URDF rendered directly in the app, while the MJCF and USD runtimes lived in isolated viewers (iframes) connected through pub/sub messages. It wasn't elegant, but it worked on desktop and even reasonably powerful phones.
platform
The platform became a full-stack Next.js and TypeScript application where people could publish, explore and fork robotics projects. I used React for the interface, Prisma with PostgreSQL for the data model, Auth.js for Google and GitHub sign-in, and Google Cloud Storage for project images and user uploads. I also built a GitHub and OnShape App integration so projects could connect back to their source repositories.
I also deployed development, staging and production environments on GCP, containerized the application and ran it on Cloud Run, stored images in Artifact Registry, and provisioned PostgreSQL in Cloud SQL. I split uploads across public and private Cloud Storage buckets, and configured bucket-level IAM and organization policies. I managed the infrastructure with Terraform, including remote state in a dedicated GCS bucket.
agentic tooling
We also explored what an AI-native robotics workflow could look like. At OpenAI's VivaLeHack, I helped build a Cursor-style agent for turning an Onshape assembly into a simulation-ready URDF.
The agent inspected the assembly, found and repaired conversion errors, and updated the robot's links and joints in a live 3D view. It could run independently, but the user could interrupt, retry a step or suggest a change at any point. We got the CAD-to-URDF loop down to roughly three minutes, built the prototype in six hours and finished as a finalist.
sunset
The useful lesson was that community pull and a viable business are not the same thing. People liked the viewer and the open-source tooling, but the broader platform did not solve a painful, recurring problem that I felt was worth building a business around. @theonlyAyo and tnkr.ai have since picked up the torch, and I hope they do very well.
I sunset the project in December 2025 to move deeper into deeptech - a path that eventually led me to FR8.