mcgyvr mcgyvr
Sign in Join

Bring your own box

Your GPU is bigger than it looks.

Lend a card. Ride the pool.

mcgyvr runs a 30B coding model on a 6 GB GTX 1660 Super. When one rig is not enough, lend your cards to the hub. Its planner splits one model across cards from many rigs, and you use it through your own endpoint.

Join Lend a rig

See the pool live

one card · 6 GB · 30B model · 25 tokens per second

Measured on one rig

6 GBGTX 1660 Super, the whole card
30Bqwen3-coder, Q4_K_M, a 17.3 GB file
25tokens per second, generating
4.3 GBof video memory (VRAM) used; the experts sit in RAM

On a rig with 32 GB of RAM. Attention on the card, 40 layers of experts in system memory. Yes, a 1660 Super. We checked twice.

How it works

Three steps, one endpoint.

  1. 01

    Run the agent

    Add your rig on the hub, then run its join command on the machine with the GPUs. Its cards join the pool while it is online. A 6 GB card counts.

    mcgyvr rig join https://mcgyvr.dev --token mhr_…

  2. 02

    Lend it

    Pick how each rig shares: the general pool, a crew you trust, or hitchhike, where riders take the spare slots of a model you already run. Change it any time.

  3. 03

    Use the endpoint

    You get an endpoint and a key that work with the OpenAI API. Point any tool at them. There is no chat on this site: you use the pool from your own tools.

The pool, right now

Everyone's cards, one machine.

1 people
0 rigs online
0 cards online
0 GB pooled VRAM
Live: updates every 10 seconds See the pool live

Pool, crews, hitchhiking

Four ways to share a rig.

You pick one for each rig, and can change it any time.

General pool the default

Your cards join everyone's. The hub runs models across them, and you use the pool through your endpoint.

Your requests run on other members' rigs. The owner of the rig that runs the start of the model can read your prompts.

Crew

Lend your cards to a group you trust. People join a crew with its invite code, and the crew's owner picks the models it runs. No credits change hands inside a crew.

Share the invite code only with people you would hand your rig to. The owner can rotate it any time.

Hitchhike

Run your own model on your own rig and let riders take its spare slots. You keep one slot, and your requests go first. You can ride on other people's models too, if you turn it on.

Whoever runs the model can read the prompts that ride on it.

Solo

Your rig runs only for you, outside the hub's sharing: it neither lends nor borrows. You can still share your setup with others.

Credits

No giving, no taking.

Lend cards, earn credits. Use the pool, spend them. The pool serves only people who lend: while one of your rigs is online and lent, you can use everyone's cards.

Your first lending earns a starter balance. Inside a crew no credits change hands: you lend, you use.

Each model on the pool's list has a price in credits per 1000 tokens, in and out.

Your endpoint

POST /v1/chat/completions

# any OpenAI API client works too; MODEL: one of GET /v1/models
curl https://mcgyvr.dev/v1/chat/completions \
  -H "Authorization: Bearer mhu_xxx" \
  -H "Content-Type: application/json" \
  -d '{"model": "MODEL",
       "messages": [{"role": "user",
         "content": "list my GPUs in bash"}]}'

Set your tool's base URL to https://mcgyvr.dev/v1 and its API key to one of your keys. GET /v1/models lists what you can use now.