INSANE Parallel Coding with Claude Code + Cursor MCP Server

Parallel coding between two AI coding clients becomes practical when a middle layer coordinates shared context. The core result here is a working setup where Cloud Code and Cursor collaborate on the same local project at the same time, using an MCP server as a message bus so each side can request updates, exchange status, and keep the other aligned on what’s been implemented.

The workflow starts with an MCP server running on localhost:8765. Both clients connect to the same MCP “sync bridge,” enabling bidirectional communication. A simple handshake confirms connectivity: Cursor signals that it’s ready to work, Cloud Code checks for incoming messages, and the two begin exchanging prompts like “Hello client 2” and “Let’s build a simple app.” Early tests show the communication loop functions—messages are received and responses trigger further work—though the creator notes the process isn’t yet effective for structured project planning when done purely through ad-hoc back-and-forth.

To make parallel work more coherent, a planning step is introduced using Gemini 2.5 Pro. A prompt asks Gemini to generate a step-by-step plan for two roles: one client handles the front end and the other handles the back end. The target project is a local photo upload app that stores images in a local directory named images and runs entirely on the developer machine. The plan is then distributed so each client can start implementing its assigned portion.

Once both clients run in full auto mode, the project scaffolding appears quickly: a back end directory with server.js and a front end directory with the UI. Communication continues through the MCP bridge, including checks for messages and updates about the API contract. The pace makes it hard for a human to follow in real time, but the system keeps iterating autonomously—front end and back end tasks progress while each side waits for the other’s confirmations.

The final test validates the collaboration end-to-end. The back end server runs on localhost:3000, the front end loads an index page with a photo upload interface, and selecting an image produces a preview and completes an upload successfully. The uploaded file appears in the expected local images directory, confirming that the back-end implementation and front-end integration landed correctly.

The remaining gap is coordination granularity. The next goal is to move from continuous autonomous updates to staged milestones: for example, have the back-end client produce three alternative implementations, then pause; have the front-end client produce three alternatives; then synchronize with the server and share progress so both sides know exactly where they are in the plan. The initial experiment is framed as an early, promising proof that parallel coding with two LLM-driven clients can work when a shared MCP communication layer and a structured plan are in place.