Your PC is the AI.

GenKaKu is a peer-to-peer AI compute network for DeepSeek Harness. Users pay nothing. Workers earn from sponsored placements and tips. No account, no orchestrator, no gatekeeper.

Live network
Read from the chain
View programs →
—
Workers registered
Solana Devnet
Settlement chain
100%
Worker share

These numbers come straight from Solana through a public RPC, read by your browser. There is no server in between and nothing is cached by us.

What is new in 0.5

Free for users. Paid by attention.

The old per-minute escrow is gone. Inference is free, a small sponsored placement covers it, and every sponsored minute is paid to the worker who served it.

Users pay nothing

No account, no subscription, no per-minute bill. The answer and the model are free.

Sponsored, not wallpaper

One clearly labeled card at the end of an answer. Each impression books a sponsored minute to that worker.

Advertisers fund their own campaign

USDC goes from the advertiser's wallet into a Solana campaign. The program pays workers; the platform never holds the money.

An agent needs permission

Off by default. An owner can let an agent ask each time, or spend within a per-request and per-day cap.

Nine languages

Both panels follow DSH's language, with a manual override that includes Korean.

How it works

Three layers, none of them ours to switch off.

The directory is a program, the connection is peer-to-peer, and the ad budget pays the worker. Every part can be inspected or replaced by anyone.

01

Discover on-chain

Workers register in the directory program with a peer ID and an optional locator. The directory is a program, not a database — anyone can read it and anyone can publish to it.

02

Connect peer-to-peer

The plugin opens a libp2p connection to the worker. If the two peers can reach each other directly, model tokens stream straight from that machine. If they cannot, the session runs through the public relay, which forwards encrypted traffic and can see its metadata — addresses, timing and volume — but not its content.

03

Show a placement, pay the worker

Each answer that shows a sponsored card books one minute to the worker that served it. Advertisers fund their own campaign on Solana; workers claim from the panel.

For workers

Serve for free. Get paid for attention.

Install the worker plugin, publish the Ollama models you already run, and set a price (0 = free). Sponsored placements pay per minute; tips go straight to your wallet. No cloud keys, no metered inference bill.

Install the worker plugin
  • Ollama only — no cloud API keys, no metered inference bill
  • Sponsored minutes and tips, with no platform cut
  • Publishing a public locator is opt-in and off by default
For users

Free. Nothing to configure.

Install the user plugin and pick a worker from the directory. The network appears as another model provider inside DSH, and inference costs you nothing.

Install the user plugin
  • No sign-up, no email, no account to lose
  • No per-minute bill: sponsored placements fund the network
  • Your key never leaves your machine
  • Tip a worker directly if you want; the transfer is yours, not ours
Verifiable

What you can check yourself.

Open programs

The directory and campaign program, and Circle's USDC, are public on Solana. Read them and every accrual yourself; the ids are on the download page.

No telemetry

Neither plugin reports usage. They talk to a public RPC for the chain, to the seed relay to find peers, and to the worker you chose. Nothing goes to us.

Receipts, not logs

Every sponsored minute is an on-chain accrual. A worker's claim is a transaction it signs; no approval from us exists.

Bounded exposure

A campaign can only promise what it funded. Agents are off by default, must ask, or run under a cap you set.

Owner keeps the keys

Export, import, rotate and identity are owner-only in every mode. A frozen wallet stops both owner and agent.

Install GenKaKu.

Two plugins: one serves this machine's Ollama models, one uses the network inside DSH. Users pay nothing; workers earn from sponsored placements and tips.

Requires DeepSeek Harness. Experimental 0.5.0-rc.8.