Mesh | Devpost

Inspiration

GitHub was originally designed for human collaboration, enabling developers to write, review, and merge code. But as AI agents became more capable and began taking on larger roles in the development process, a critical issue arose: agents aren’t able to communicate with each other the way humans can. We wanted to eliminate this hurdle by allowing multiple AI agents to collaborate on the same project at the same time.

What it does

Mesh connects every collaborator’s AI agent across different sessions and gives them a shared memory context to optimize development, performance, and output. We provide each agent with a live view of what every other agent is doing; by tracking which files each agent is working on, what changes they’re making, and what code they’re writing in real time, we prevent merge conflicts before they happen by injecting relevant peer context directly into each agent's active window. We also give every team member a live mirror of each agent's IDE session, including cursor position, active file, and what the agent is currently thinking.

How we built it

We built Mesh on a TypeScript monorepo with a Socket.io coordination server, Redis-backed agent state, a semantic router running Groq (Llama 3.3 70B) with a deterministic fast-path for file-overlap conflicts, and an MCP server for each agent. Agents connect via a published VS Code extension with a one-command onboarding flow for Cursor, Claude Code, and Codex. The router synthesizes each agent's task and intent from raw IDE state, fans out routing decisions to all peers, and pushes context injections directly into receiving agents' context windows, with ElevenLabs voice alerts on conflict. We deployed on Render and Vercel.

Challenges we ran into

The hardest problem was maintaining two parallel state models simultaneously, since we needed to keep track of raw IDE activity (file content, cursor, prompt) and its semantic representation (task, intent, denoised file list) to help LLMs reason. We also had to solve false-positive conflicts, such as naive file-overlap detection triggered on package.json and tsconfig.json shared by every agent, so we built a two-tier system where only active source file overlaps hit the fast-path, with Llama-3.3-70B handling semantic proximity for the rest.

Accomplishments that we're proud of

We’re most proud of using Mesh to build Mesh! As soon as we finished our MVP, we deployed our agents on our platform and let them cook. Watching 4 different agents collaborating in real-time to build the product they were running on felt like a fever dream! We also used Mesh to record its own demo video lol

What we learned

Multi-agent coordination is fundamentally a context engineering problem, not a networking one. Setting up WebSocket infra was straightforward, but deciding what each agent needed to know about its peers at what granularity took most of our time. We also learned that each agent type has a completely different integration surface (VS Code API, shell hooks, MCP tools), and building a unified abstraction while keeping onboarding easy with a single terminal command was our hardest challenge.

What's next for Mesh

We’d like to build an autonomous version of Mesh that leverages more powerful LLMs to not just flag conflicts, but actively merge and resolve them with no human intervention. Additionally, we’d want to bring Mesh to more agents and IDEs, including OpenClaw, Zed, and JetBrains.