I built GPUMesh with Claude Code - my AI agents can now run on my friend's idle GPU
A developer created GPUMesh, a solution for utilizing idle GPUs, particularly for AI agent tasks. The inspiration came from the common issue of a personal GPU being occupied while a friend's NVIDIA GPU remained unused. The developer successfully tested the complete flow on an RTX 5060, which involved pairing, connecting, sharing, launching a remote Docker job, and running nvidia-smi within a remote CUDA container. The creator is seeking feedback on potential issues and desired future features.
Time & source
Times shown in UTC
Display time zone: UTC
Local time zone unavailable; showing UTC.
PublishedOffset at this time: UTC+0Sep 13, 2026, 07:35 UTC
IngestedOffset at this time: UTC+0Sep 13, 2026, 11:01 UTC
- Published
- Sep 13, 2026, 07:35
- Ingested
- Sep 13, 2026, 11:01
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
While building AI agents, I kept running into a simple problem: my GPU would be busy training/running something, while my friend's NVIDIA GPU was sitting completely idle.
I wanted a simple way to use that spare compute without turning it into a full cloud infrastructure project.
So I built GPUMesh, an open-source P2P GPU sharing tool.
The idea is that trusted machines can pair with each other and run Docker GPU workloads remotely:
gpumesh share → share a GPU gpumesh pair → establish a trusted connection gpumesh run --peer ... → run a GPU job on that machine
The workload actually runs in a Docker container on the provider's machine, next to their GPU. GPUMesh handles the peer connection, permissions, job execution, and logs.
I've been using it to:
- Run AI agent workloads on another machine
- Train models when my own GPU is occupied
- Use spare GPU capacity from friends
- Run CUDA/Docker workloads without manually setting up another server
How Claude Code helped
I built a large part of the project using Claude Code.
I used Claude Code to help design and implement the CLI and Rust workspace, work through the P2P networking layer, container/job execution flow, pairing and access-control logic, GPU detection, error handling, and testing.
One thing I particularly liked was being able to give Claude Code the higher-level requirement — essentially “I want two trusted machines to securely discover each other and let one submit a Docker GPU job to the other” — and then iterate through the networking, CLI, and execution details with it.
Claude Code was also useful for debugging the messy integration points between networking, Docker, NVIDIA Container Toolkit, and the CLI.
The project is free and open source. You can run it yourself and use your own machines/GPUs.
I've tested the complete flow on an RTX 5060: pairing → connecting → sharing → launching a remote Docker job → running nvidia-smi inside the remote CUDA container.
arjun988/GPU-Share
I'm curious what people here would build if their AI agents could access a few trusted GPUs across their machines.
If you try it, I'd be interested in what breaks or what you'd want it to support next. A ⭐ on the repo is also appreciated if you find the project useful.