A MacBook Air prompts, a Windows PC does the work: remote MCP for circuit design

The problem: a big program, a small laptop

Cadence Allegro X is software for designing multi-layer printed circuit boards (PCBs), the boards that hold the chips and wiring inside electronics. As dev.to reported, the official guidelines ask for at least 50 GB of free disk space and up to 64 GB of RAM for intense routing jobs.

The author's daily machine is a MacBook Air M2 with a 256 GB drive and 16 GB of RAM. Running Windows inside a virtual machine, for example with Parallels Desktop, would split that memory between two operating systems and make both sluggish. The virtual disk alone would take more than 120 GB, leaving the SSD close to full.

Split the thinking from the doing

The guide's answer is to separate where you give instructions from where the work runs. The Mac holds the conversation with the AI agent. A separate Windows machine, described as an Infinix GT Book with a Core i9 chip, 32 GB of RAM and a discrete RTX graphics card, does the heavy lifting.

The glue is the Model Context Protocol (MCP), a standard that lets an AI model call outside tools. Usually an MCP tool runs as a small process on the same computer. Here the guide uses a remote setup instead, reached over the network with HTTP and Server-Sent Events (SSE), a way for a server to keep streaming messages to a client.

The Mac side uses Google Antigravity as the agent. It acts as the human-in-the-loop client, meaning the person types the request and checks the results. The Windows side runs a server called winremote-mcp.

How the setup works, step by step

On the Windows machine, you install the bridge with pip install winremote-mcp, then start it with a host setting of 0.0.0.0 and port 8090. You also note the machine's local IPv4 address; the guide uses 192.168.1.50 as an example.

On the Mac, you edit the agent's configuration file, typically named mcp-config.json or agent.json. There you register a server entry that points to the Windows machine's address on port 8090, at the path /sse, and mark it as remote.

After that, a request flows in four stages:

• You type a design request on the Mac, such as importing footprints from a library and running a placement script on layer 3.
• The AI turns it into an ordered list of steps.
• Those steps travel across the local network as standard tool-call messages.
• The Windows server receives them, maps them to shell tools or Cadence command-line programs, and runs them on the stronger hardware.

Two practical catches

The first is file paths. Because the work happens on Windows, your prompts must name files the Windows way, such as C:\Cadence\Projects\, and not the way a Mac does. An agent told to open a Mac-style path will point at a location that does not exist on the worker.

The second is security, and the guide is blunt about it. Basic remote MCP tools typically ship without strong built-in login checks. A server that can run shell commands for anyone who reaches it is a risk, because it could allow remote code execution, meaning a stranger running commands on your machine.

The advice is to keep the whole arrangement on a trusted local router or behind an encrypted VPN link, so only your own devices can reach the worker.

What this does and doesn't show

The write-up is a how-to, not a test report. It gives no timings, no sample boards and no account of how well the AI handled real Allegro X jobs. It shows that the plumbing can be arranged, not that the results are good.

Even so, the pattern reaches beyond circuit design. Any Windows-only or resource-hungry desktop program could in principle be controlled this way from a lighter machine. The guide also shows what the decision involves: you save local disk and memory, but you take on a network service that needs careful protection.