gpuroute

This is a public snapshot. Install gpuroute and run gpuroute ui for your live dashboard.

Start^{(3 steps)}

One command to connect your agents, one to share a spare machine, then just ask.

(01)

Connect

Checks this machine and adds gpuroute to Claude Code and Codex, or prints the config for Cursor.

uvx --from "git+https://github.com/sissississi-013/gpuroute#subdirectory=gpuroute" gpuroute init

(02)

Pool

Run gpuroute hub once to get your pool code. Then anyone, anywhere (a friend’s gaming PC, a lab box, a teammate’s laptop) joins with one command. It connects outbound, so there’s no network setup. It only takes jobs while idle, and the owner sees each one.

gpuroute hubyou, once: prints the pool code

uvx --from "git+https://github.com/sissississi-013/gpuroute#subdirectory=gpuroute" gpuroute share --pool "<pool code>" --name <name>

(03)

Ask

Your agent calls compute_map, run_workload, quote_job and launch_job itself.

“What compute can you use right now?”
“Embed my notes on a GPU. Don’t hog my laptop.”
“Fine-tune Qwen2.5-0.5B on my data for under $1.”

Map^{(land/sea/orbit)}

Where the world’s compute physically is, and how much of it your agent can actually use. Agents query the same data with world_compute.

(ORBITAL MARKET: COMPUTE + RELAY, NEXT 6 HOURS)

Agents buy GPU time aboard satellites while they’re sunlit, plus ground-station passes to downlink results: orbital_offers → book_orbital. Windows and passes come from the real orbits. Operator listings and prices are mock, and jobs are emulated on ground GPUs.

Sources: ATLAS Global Data Center Map (6,131 facilities with coordinates) · Epoch AI Frontier Data Centers (CC BY 4.0; geocoded with OpenStreetMap Nominatim) · TeleGeography Submarine Cable Map · CelesTrak orbital elements (positions computed live with SGP4) · underwater sites from public reporting, locations approximate.

Compute

Everything your agents can reach, and how busy it is right now. Agents see the same list through compute_map.

(SHARE A MACHINE)

Run this on any Mac or Linux box with a GPU. Jobs only start while it’s idle, each is capped at 30 minutes, and the owner sees every job.

uvx --from "git+https://github.com/sissississi-013/gpuroute#subdirectory=gpuroute" gpuroute share

gpuroute join name@ip:7400#tokenpaste what it prints

Jobs

Anything your agents ran: training, embeddings, transcription, benchmarks. Outputs come back to this machine.

gpuroute run my_script.pydeps are read from its imports

Spend

Agents hold an allowance, never your credentials. Paid jobs above the threshold wait for you, and the budget becomes the provider’s hard timeout.

(PROVIDERS)

(PRICE A FINE-TUNE)

Start(3 steps)

Map(land/sea/orbit)

Compute

Jobs

Spend

Log

Start^{(3 steps)}

Map^{(land/sea/orbit)}