A Client Worth Watching: Upfreq Robotics
We're glad to have Upfreq Robotics on our roster of clients. Upfreq is a desktop app, Upfreq Studio, where a team of AI agents builds and tests software for your robot in simulation, from the first line of code to a merged pull request. The company was founded by robotics software engineers around one idea: raise the frequency at which robotics teams build, test and ship. It overlaps closely with our own mission: software where agents are the primary interface, not a feature bolted on afterward.
What Upfreq Does
Robotics software is hard to iterate on because the code, the robot and the world it moves in all have to work together. Upfreq keeps the robot in the loop, so agents know the robot and not just the source files. The workflow has four steps:
- Import a robot — drop in the robot's model (URDF or xacro). Upfreq checks every file, builds a simulation-ready digital twin and records the robot's ROS 2 interface so every agent knows what the robot listens to and publishes.
- Create a sim setup — put the robot in a world (a built-in warehouse, hospital, office or test track, or your own USD environment), place objects, pick a start pose, and check it fits on your GPU.
- Start working on a feature — describe what the robot should do in a line. The agent reads the robot's interface, writes the code and tests, and runs it in the live simulation (NVIDIA Isaac Sim).
- Review, test, merge — agents open the pull request, answer review comments and fix CI when it breaks. You read the change and merge.
Several agents can work side by side on one GPU, each in its own workspace with its own branch and ROS domain, so features don't step on each other. Upfreq supports mobile robots, robot arms, drones, self-driving cars and humanoids. It runs on your own GPU (Basic, from $10 per user per month) or on a dedicated Upfreq GPU (Pro), with custom Enterprise plans, and is currently in early access. See the Upfreq docs for the full walkthrough.
Why This Is an AI-Native Problem, Not Just an Automation One
What drew us to Upfreq is that it isn't a code autocomplete with a robot attached. The agent plans the work, writes the code, writes the tests, runs them against a live simulation and reacts to what happens, the way a robotics engineer would. That's the same architectural bet we make with every client: treat the agent as the primary actor in the system, and design the surrounding infrastructure — data access, tool calls, guardrails — around it from day one, rather than retrofitting AI onto a traditional pipeline.
Our Role: Web Presence, MCP Servers, and Agent Harness Engineering
Our engagement with Upfreq started with their web presence — design and build, end to end. Since then, we've stayed on as a partner for agent-native engineering: building MCP servers that expose robot and simulation tooling as callable tools for AI agents, and supporting the agent harness work — the orchestration layer that governs how an agent calls tools, holds state across a session, and hands control back to a human when a result needs review.
This is the same practice we lean on across our client work — see MCP servers and agent harnesses for how we think about it in general: MCP server creation, so a client's internal systems become first-class, callable capabilities for Claude and other AI agents, and agent harness builds, the execution and memory layer that turns a capable model into a reliable, production-grade agent. For a company building AI engineers for robotics, getting that layer right matters as much as the simulation itself.
Why This Partnership Matters
Robotics is one of the clearer proof points for agent-native software: the cost of a slow, manual build-test loop is measured in weeks of engineering time and delayed hardware releases, not just inconvenience. Moving that loop into simulation, with agents doing the iteration and a human reviewing the pull request, is a preview of what happens across other industries once agents are treated as core infrastructure instead of an add-on. It's why we keep building this muscle: MCP servers and agent harnesses aren't a niche offering for us, they're the backbone of every agent-native system we ship, including our own products like Physiolaxy.
Frequently Asked Questions
What does Upfreq Robotics build?
Upfreq Robotics builds Upfreq, a desktop app where a team of AI agents builds and tests software for your robot in simulation, from the first line of code to a merged pull request. It is in early access.
Which robots does Upfreq work with?
Mobile robots, robot arms, drones, self-driving cars and humanoids. If your robot has a model (URDF or xacro), Upfreq can simulate it.
What did Vantaverse build for Upfreq Robotics?
We designed and built Upfreq's web presence, and continue as a partner for MCP server development and agent harness work.
What is an MCP server, and why does it matter for a robotics AI company?
An MCP (Model Context Protocol) server exposes a system's tools and data as structured, callable capabilities an AI agent can use directly. For a robot software agent, that means it can read the robot's interface, trigger simulation runs, or pull results as native actions instead of brittle, one-off integrations.
Conclusion
Upfreq Robotics is a good example of what we mean by agent-native software: agents that work on a real, expensive engineering bottleneck, with the robot in the loop and a human reviewing the result. We're glad to have built their web presence, and happy to keep building the MCP servers and agent harness infrastructure beneath it as our partnership continues.