Sipeed, the Shenzhen-based hardware company known for RISC-V boards and edge AI gadgets, launched a Kickstarter on July 1 for the NanoKVM-Go, a USB-C KVM the size of a watch. The device does what any KVM does — control a computer’s display, keyboard, and mouse remotely — but it wraps that capability in a package designed for AI agents. At $59 early-bird pricing, it is a bet that the next generation of AI tools will need physical access to screens, not just APIs.
The NanoKVM-Go connects to any device with USB-C DisplayPort Alt Mode — iPhones from the 15 onward, Android phones, MacBooks, Steam Decks, Windows laptops. A single cable carries video, audio, HID emulation, disk emulation, and a virtual NIC. The host connects over WiFi 6. The device itself runs on an Axera Tech AX630C system-on-chip, with a dual-core Cortex-A53 CPU, a 3.2 TOPS NPU, and 256 MB of LPDDR4x RAM on the base model.
The Go+ variant doubles the RAM to 512 MB and bumps storage from 16 GB to 64 GB eMMC. That extra capacity is for what Sipeed calls “full screen memory” — a local screenshot database searchable by natural language, functionally identical to Microsoft’s Recall feature. Sipeed claims the Go+ can hold up to 180 days of Recall data on its 64 GB of storage.
The AI features are what make this device worth watching. Sipeed says it is the first “AI-Native KVM” — all KVM functions are exposed as an MCP (Model Context Protocol) server. An AI agent can use the MCP server to see the screen, move the mouse, type text, and click buttons. The device becomes a hardware-level Computer Use Agent peripheral, paired with any AI agent that speaks the protocol. Sipeed also supports its own lightweight agent, PicoClaw, for local inference on the device’s 3.2 TOPS NPU.
This is a fundamentally different approach from the software-only computer-use agents shipping today. Anthropic’s computer-use feature in Claude operates through screenshots and simulated mouse movements on the host machine. OpenAI’s Operator does the same. Both run entirely in software, inside a virtualized browser or a desktop application. They are brittle — they break when a UI changes, when a dialog box steals focus, when the system lags.
The NanoKVM-Go operates at the hardware layer. It emulates a USB keyboard and mouse. It captures video frames from the DP Alt Mode signal. It does not care what operating system the target runs, what browser it uses, or whether the UI is a web app or a native application. The agent sees exactly what a human would see on the screen, and it interacts through the same channels a human would use. There is no abstraction layer to break.
The latency numbers are real enough for practical use. Sipeed quotes 60 ms at 1080p60, 80 ms at 2K60, and 100 ms at 4K30. That is slower than a direct display connection but faster than most remote desktop protocols over the internet. For an AI agent clicking buttons and reading text, 100 ms of latency is negligible.
The Kickstarter has already blown past its $6,374 US funding target. The “Early Bird” Go is $59, the Go+ is $79. Retail prices are set at $89 and $129 after the campaign ends in August 2026. Sipeed has shipped KVM products before — the original NanoKVM and the NanoKVM Pro 4K — so the risk of non-delivery is lower than for a typical crowdfunding project.
The closest competitor is the GL.iNet Comet Q (GL-RMQ1), also on Kickstarter, which has raised over $1 million. The Comet Q is larger, supports lower resolutions, and lacks the NPU and MCP server. Sipeed’s advantages are the smaller form factor, the 4K support, the larger built-in storage, the FingerBot support for physical robot control, and the AI features.
The FingerBot support is worth noting. The auxiliary USB-C port on the NanoKVM-Go can power and control a FingerBot — a small robotic finger that presses physical buttons and switches. This extends the agent’s reach from digital screens to physical controls. An AI agent can navigate a touchscreen, then reach over and press a hardware power button. This is the kind of integration that makes sense only when you think of the device as an AI peripheral first and a KVM second.
The GitHub repository for the NanoKVM-Go exists but contains only a GPLv3 license file. The wiki and documentation are empty. Sipeed says they will be populated after the campaign. The open-source license is a meaningful signal — it means the MCP server implementation and the agent integration code will be inspectable and forkable.
The implications for AI builders are straightforward. Software-only computer-use agents are hitting a wall. They work well in controlled environments — a headless browser, a clean desktop — but fail in the wild. A hardware KVM with an MCP server removes the abstraction penalty. It gives the agent the same interface a human has: a screen, a keyboard, a mouse. If you are building an agent that needs to interact with legacy software, with physical machines, with any system that does not expose an API, the NanoKVM-Go is the cheapest way to give that agent eyes and hands.
The bigger picture is about the shape of AI infrastructure. The dominant paradigm today is cloud-to-cloud: an API call to a model hosted on a GPU cluster, returning text or structured data. The NanoKVM-Go represents the opposite pole: a local device with a small NPU, connected to a physical screen, controlled by an agent that may itself run on a cloud model. It is a bridge between the digital and physical worlds, priced at $59.
Whether the market for hardware-level computer-use agents is large enough to sustain a product category is an open question. But Sipeed is not asking for much — $59 for a device that also works as a perfectly good KVM. The AI features are a differentiator, not the whole product. That is a sensible hedge for a hardware company selling into a market that may or may not materialize at scale.
The $59 early-bird price will not last. After August, the NanoKVM-Go costs $89, and the Go+ costs $129. At those prices, the device still undercuts every other 4K KVM on the market, with or without AI features.