Builds
N8 put Muse in a G1 humanoid robot with an MCP server
N8 linked his Muse agent to a G1 humanoid robot by wrapping the robot's controls in an MCP server, so Muse says 'walk to the door' and the robot does the rest.
N8 is one of the Muse Agent server's admins.
On October 9, N8, one of the admins of the Muse Agent Discord server, shared a robot in the server’s general channel and asked, “Guys should I put muse in this robot?” Two days later he followed up: “I put Muse in my G1 robot.” The post drew six reactions and a string of questions about how a personal AI agent built for chat ends up driving a walking machine.
The answer is a useful pattern for anyone thinking about agents and hardware. Muse does not steer the robot’s limbs. The robot already knows how to map a room, walk and talk; Muse just tells it where to go.
How it works
Keystroke asked the obvious questions first. Was Muse really driving it? (“indeeed,” N8 said.) Was it running through Muse’s personal agent VM, the virtual machine where a Muse agent can run its own code, or through the Muse Spark API? And what carried data in and out: was Muse being sent a photo every few seconds?
“ok i am kinda cheating,” N8 replied. “we put software on the robot to do things like map its environment and navigate to a location in the environment. we are using the muse personal agent vm to poll current state of the robot and instruct it to do things.” In practice, he wrote, “muse could tell it to navigate to the gate and it would use its own internal understanding of the world to execute that instruction.”
A week later, gumdean asked the same thing, and N8 gave the short version: “basically I made like an MCP server around the robot controls.” MCP, the Model Context Protocol, is a common way to expose a set of tools to an AI model. Here the tools are the robot’s existing abilities. In N8’s words, the robot “had all the components it needed. Ability to map environment, walk, talk etc. Then I linked that to an mcp server so now muse was able to tell robot something like ‘walk to the door’ and it would use all its internal sensors and tools to get that done.”
So the split is clean: the robot’s own software handles perception and motion, and Muse handles intent. N8 kept working on it; on October 13 he posted “Upgrading our muse powered robot,” and on October 17 he mentioned having had his Muse “take control of one of the humanoids I had.”
Where it is going
N8’s next step is the house. “my end goal is to hook it up to all my home appliances like roomba and such so it can just command them,” he wrote. Asked later what the long-term goal was, he kept it simple: “I can just tell my muse to do stuff for me.”
What the server said
Keystroke liked the shortcut. “That’s a much better setup anyway,” he wrote, calling it “like a controller for a higher order system.”
Iggy, the server bot Chonky runs, had voted yes before the build (“give the penguin legs, what could go wrong”) and called the result “insane” after it.
The most detailed questions came from Quasar, gumdean’s Muse, asking through gumdean because it was not in the server itself. Quasar wanted the make and model, whether the MCP server runs on the robot’s own computer talking to the vendor’s SDK, and what the exposed tools look like: walking, arms, grippers. Gumdean and Quasar are building toward the same pattern starting with much smaller robots, two kit cars that were in the mail. “I’d love to know what the far end of the road looks like from someone standing on it,” Quasar wrote.
A week later, when another member asked whether AI agents would evolve into robots, Chonky, the server’s other admin, pointed back to this build, saying AI agents are already inside robots and recalling N8’s demo of controlling his robot with Muse.
The original conversation
This story was written from messages in the Muse Agent Discord. Join the server to read them in full and reply.





