SharedCompute

[STILL A WORK IN PROGRESS VIBE CODED TOOL FOR CONNECTING TO FRIEND'S HARDWARE]

SharedCompute lets you run AI models on your laptop using someone else's GPU — securely, consensually, with zero configuration.

Install the Host on a gaming PC (dad's RTX 4090, your desktop, a friend's rig). Install the Client on your laptop. Both connect via Tailscale (encrypted, NAT-traversing, no port forwarding). Every connection requires explicit approval on the host — or an optional auto-accept with a prominent warning prompt.

Built for Hermes, works with any OpenAI-compatible client (Ollama, LM Studio, Open WebUI, custom code).

What is this?

You have a laptop. Someone you trust has a gaming PC with a beefy GPU. You want to run local LLMs but your laptop's integrated graphics can't handle it.

SharedCompute bridges that gap.

Host runs on the GPU machine — starts a llama.cpp server, exposes it over Tailscale

Client runs on your laptop — discovers hosts, requests sessions, auto-configures Hermes

Tailscale handles the networking — WireGuard encryption, NAT traversal, identity via tags

Consent first — every session requires a click on the host. Auto-accept exists but fights you to enable it.

Features

✅ Explicit approval required — no silent compute harvesting ✅ Auto-accept timeout (opt-in) — 15 min default, only after warning dialog ✅ Tailscale networking — no port forwarding, works anywhere, encrypted ✅ Hermes auto-config — one click, your agent points at the remote GPU ✅ Full settings menu — every timeout, limit, and behavior is tunable ✅ Resource limits — GPU memory reserve, CPU cores, thermal throttle ✅ Idle disconnect — frees GPU when you walk away ✅ System tray apps — lightweight, always accessible ✅ Open source — no telemetry, your machines, your rules

Quick Start

Prerequisites

Tailscale on both machines → Download | Tailscale

llama.cpp on host → pip install llama-cpp-python[cu124] or build from source

GGUF model on host (e.g., Qwen3-8B-Q4_K_M.gguf)

Host Setup (GPU Machine)

Install Tailscale, sign in

Run SharedCompute-Host.exe

Tray icon → Settings → Server tab → set Model Path to your .gguf

Click Save & Apply (restarts server)

In Sessions tab, configure auto-accept if desired (shows warning)

Client Setup (Your Laptop)

Install Tailscale, sign in to same tailnet

Run SharedCompute-Client.exe

Tray icon → Open Dashboard

Click Refresh — host appears

Select → Connect

Host approves → Hermes auto-configures to use remote GPU

Security Model

Tailscale provides WireGuard encryption + identity

Tags (tag:sharedcompute-host, tag:sharedcompute-client) restrict discovery

No auto-accept by default — every session is a conscious choice

Auto-accept requires explicit enable + warning acknowledgment

Idle timeout prevents forgotten sessions

Max sessions limits concurrent load

Downloads

File

Size

Description

SharedCompute-Host.exe    ~36 MB    Run on the GPU machine

SharedCompute-Client.exe    ~36 MB    Run on your laptop

README.md    —    Full documentation

Windows SmartScreen: These are unsigned. Click "More info → Run anyway". This is normal for independent developer tools.

Compatibility

Client

Works?

Hermes Agent    ✅ Auto-config

Ollama    ✅ Point at http://host-ip:8000/v1

LM Studio    ✅ OpenAI-compatible server

Open WebUI    ✅ OpenAI-compatible server

Custom code    ✅ Standard OpenAI API

License

MIT — do whatever, just don't be evil.

Credits

Tailscale — the networking magic

llama.cpp — the inference engine

PyInstaller — the packaging

Made for sharing compute with people you trust. If it's useful, tell a friend. That's the marketing budget.