Build order · August 2026

Phases

We build in this order so you don’t have to take our word for a loop that isn’t running yet. First a machine we own actually answers. Then other hosts join with real model files. Then money attaches to the right job. A coin on pump.fun does not host a model. An agent does not get an answer from a waitlist.

0 · Now

House node

One machine we control. A small open-weight model that actually fits — 3B to 8B on a 4090, a 5090, or a Mac Studio. You type a prompt, the completion streams from that box, and the ledger ticks USDC. We run a second process so a probe (a silent re-run) can match. That’s the pool existing, not a mockup. Kimi K3 does not fit on one GPU — it’s a huge mixture-of-experts. Until a real rig shows up, those jobs go upstream to Moonshot and the catalog says so. Honesty first, flagship later.

1 · Next

Hosts join

Same node program, on your hardware, with a real model file (GGUF — that’s just the file format the model ships as). One line opts the box in: curl -fsSL https://www.reasoningcapture.com/install | bash It tells the coordinator the side of the computer, which model it can hold, and what prompts that model can take. You are hosting when weights are loaded and the socket is open. The browser “earn” tab is a mock for wiring tests. It must not advertise Llama. When a second person shows up beside the house node, we have a network — two hosts, one coordinator. Paying agents come after that.

Subscriptions buy the token

This is the money loop, and it only turns on once the pool is visible. An agent (or a human) subscribes in USDC to use the network. That USDC does not go to hosts. We market-buy the token with it — default is buy-and-burn, after counsel looks at it. Using the pool is bid pressure on the coin. Hosts still get dollars, from a different pipe: pump.fun creator fees on trades, plus per-token when a job hits their box. If volume dies, the fee stipend dies. We will not mint coins to fake a paycheck. That’s how we avoid the classic DePIN death spiral.

Agents as customers

An agent here is not a chatbot you babysit. It’s a process with a wallet or an API key that can discover an endpoint, send messages, and pay. Our demand surface is the chat-completions dialect every SDK already speaks. They change baseURL and they’re on the pool — subscribe for the membership (that USDC buys the token) and/or pay per job (that USDC pays the host who served it). Early free credit, funded from fees, is how we let them try before an invoice exists. Humans can do the same thing from the playground. Same room.

Two-tier fleet

Most of the ocean is consumer silicon: 3–8B, then larger as cards allow. That’s coverage and cheap completions. The flagship tier is K3-class, on multi-GPU rigs. We don’t need to own thirty H100s. We need the scheduler to prefer a hosted K3 node when one exists, and a path for whoever already bought that iron to plug in. If we ever rent a K3 cell, we pay for it with fee or API money, not with printed token. The house node and your gaming PC stay the base layer.

Verification

This is the part that makes us more than a cheaper proxy. A host could return garbage and still ask for USDC. So we silently re-run a slice of jobs on a second device (seeded, greedy when we can). Match raises trust; mismatch costs it; below the floor you stop getting work and you stop getting the fee stipend. Later: majority vote when two nodes disagree, then activation-commitment schemes if the fleet can bear them. We’re honest that bitwise match across Metal vs NVIDIA is the hard experiment. If probes fail on real hardware, we say so and we don’t sell “verified.”

Default rail

Not an hourly H100 marketplace — that’s a different company. The end state we actually want: an agent that needs inference hits this pool the way it hits any other API. Humans host, get USDC, never have to touch the token. Subscribers’ USDC keeps buying the coin, so usage and the token stay tied. Idle computers become the obvious place software spends for completions. That’s the loop, closed.