A skill is not a worker

This update was not a recap of what shipped. Nils wanted to give a clearer picture of where we're going: how the real world web actually gets used, and why the next step is a harness for physical AI.

Two pieces still sit at the center of the company. The bigger project is making the physical world accessible to AI — an internet of sensors, an internet of actuators, and an internet of spaces. The thing customers already run is the retail copilot. Today's argument was how those two meet.

The Auki SDK is how you plug into the real world web

We've been building a revamped, fully open-source Auki SDK. It is not the old posemesh repo. It is still experimental. Some of it is already in production. The mature Posemesh SDK is what we run with customers today; more of the experimental stack keeps landing there.

The SDK is aimed at three problems that make robots miserable to deploy:

• The environment. Every site is different, and the important ones keep changing. Today's default is that each robot perceives the space on its own, with no shared state. We're building networked, composable scene graphs — collaborative spatio-semantic models of the place the robots work in.
• The embodiments. Not the metal. The sensors and actuators that pierce into the world: cameras, mics, speakers, displays, grippers. The landscape used to be phones. Now every machine exposes itself differently, and fleets have no common way to publish, consume, or compose capabilities. The SDK treats the fleet as a dynamic catalog of sensors and actuators, so an agent can ask for a data stream or an actuator instead of "this specific robot" — and stitch just-in-time pipelines, even virtual robots, out of pieces that already exist.
• The network. Every deployment has its own topology, servers, firewalls, and security policy. The SDK's answer is effortless, secure peer-to-peer networking that works across embodiments and across applications.

In Nils's words: the Auki SDK is how you build harnesses for physical AI. Physical context management, tools for collaborative spatial reasoning, and networked composable embodiments.

Cursor for code. A digital twin for space.

A harness, as we mean it, is what Cursor is for a coding model: context management plus tools. Talking to ChatGPT in a browser is weaker because the model can't see the codebase or act in it. Cursor can.

Physical space doesn't come with a repo. A lot of the context is missing — the world isn't digital. If we want frontier models (ChatGPT, Claude, Gemini) to be useful on the front line, we have to get physical context and business context into a live, collaboratively editable digital twin. That's what context looks like for a store, a warehouse, a home.

We want to connect the frontier to the front line.

World models don't know what to do next

Robots are good at repetitive work. They struggle in high-mix, low-volume environments: many kinds of tasks, none of them running for long, priorities that shift. World models and VLAs can predict physical outcomes. They cannot predict the economic impact of a choice. So they still need a human to point.

The household example: more chores than hours, more chores than battery, and a 30-minute window before you walk in the door. Empty the dishwasher, clean the bathroom sink, or get the laundry off the couch so you can sit down. A world model does not know which of those is worth the remaining charge. If the laundry takes 40 minutes and you have 30 minutes of battery, maybe the robot empties the dishwasher for 15 minutes and goes to dock. Robots are nowhere close to that kind of call today, because they don't have that kind of context.

Being able to learn a new capability in the wild is not the same as being a self-directed and autonomously productive worker.

The path we think is real: a local domain model — a live digital twin, plus agentic reasoning, plus tools to simulate the future of that twin so the system can pick the next job. Cactus is the first early version of that. It already builds and maintains a digital twin and reasons over past, present, and (increasingly) future so work can be directed across a site's fleet.

We wouldn't be able to build that harness without the Auki SDK. The most useful version of it is networked. That's how you get back to the real world web.

What Cactus is already being asked

In retail, Cactus is the harness in front of a frontier model. You can already ask things like: which area of the store is the least performant by square meter, why, and what would you change. The answers are specific enough to run a store on, not a slide on.

The next move is plugging robots into that same brain — not only as named machines, but as composable capabilities. Galbot, RealMan, and BracketBot are in that picture. A new BracketBot showed up in the lab today. It has arms.

We hosted Robots & Beers at the lab this afternoon — about 100 people from the robotics community, live demos from people building on Auki. We ran out of beer and food.

Nils is also putting this argument into an article on X this weekend: what a general-purpose robot actually is. Keep an eye out.

We do these every week. After the broadcast we hang out off the record on Discord: discord.gg/auki.

Watch the whole update on X.

About Auki

Auki is making the physical world accessible to AI by building the real world web: a way for robots and digital devices like smart glasses and phones to browse, navigate, and search physical locations.

70% of the world economy is still tied to physical locations and labor, so making the physical world accessible to AI represents a 3X increase in the TAM of AI in general. Auki's goal is to become the decentralized nervous system of AI in the physical world, providing collaborative spatial reasoning for the next 100bn devices on Earth and beyond.

X | Discord | LinkedIn | YouTube | Whitepaper | auki.com