FAQ

Basics

What is Reasoning Capture?

A network of home and office computers running open-weight models. If you host, the model lives on your machine, the answer is written on your machine, and you get paid in USDC. Buyers are people and agents — software that already talks to ChatGPT-style APIs. We sit in the middle: pick a host, meter the job, spot-check a slice, settle the dollars. You’re not reselling a lab’s API. There’s also a token; you can ignore it and still host.

What do you mean by “agent”?

Software that can call an API and pay without a person clicking each time — a coding assistant, a scraper, a pipeline, an agent harness. Not the program you install to host. That one we call the node program. Agent = customer. Host = earner. Easy to mix up; we mixed it up on earlier drafts.

What is a "job"?

One request: text goes in ("extract the total from this receipt…"), text comes back ("$482.19"). Each job runs on exactly one device. Large workloads are thousands of jobs spread across many devices in parallel. Every job is metered — who asked, which device served it, how many tokens, what it cost.

What’s the terminal?

A chat window on this site. That’s our first node. Join the waitlist on that page if you want to host, or get a key.

Is this live?

Not yet, not for everyone. Join the chat and leave an email — or the Chat button — and we will send an invite when we have room.

Earning

How do I start earning?

Follow How to join. After an invite, one line on a Mac or Linux box that can stay awake: curl -fsSL https://www.reasoningcapture.com/install | bash That reads the machine (OS, RAM, CPU), picks a catalog model it can hold, and opts it into the roster — what prompts it can take, so the coordinator can assign work. Serving paid jobs still needs the public coordinator and a real model file (GGUF) on disk. The browser “earn” tab is only a wiring test; it does not run named models.

What can my computer run?

These are the options. We only list models that actually fit. The install line reads RAM and CPU and records which option you are. A box that sleeps in a bag earns nothing. Awake, model loaded, spot-checks passing — that’s a host.

Each model is a different job, not a second name for the same 7B. You pick one file (GGUF) and stay with it.

Mac, Windows, or Linux — if it can stay awake, it’s in the same menu. The one-liner today is Mac and Linux; Windows through WSL. Phones later, for the small class. The coordinator is up. The install line will start the node program when it can reach it. Serving a named model still needs a real model file (GGUF) on disk.

I have an old MacBook Air. What can I run?

Same menu as any other computer. Intel Air, 4 GB — sit this one out. 8 GB Air — small host: Llama 3.2 3B or Phi-4 mini; it will warm up, leave it plugged in. 16 GB Air — everyday: Qwen 2.5 7B, Hermes 3, or the DeepSeek distill. Kimi does not fit on one laptop. An Air in a bag is worse than a cheap Mini on a shelf.

What is my computer actually doing?

It loads the model you picked and waits for work. When a request is sent to you, your computer writes the answer itself. Text in, text out. You are not forwarding the question to a cloud lab or anyone else.

Do I need an AI subscription or an API key from anyone?

No. Open-weight models are free for anyone to download and run. You're not reselling access to an AI company — you're selling your computer's time.

How safe is this?

Jobs are text in, text out through a model you installed. A buyer cannot run a program on your machine. There is no remote desktop and no shell. You pick the model. You can quit; work in flight moves to someone else.

The one-liner is a roster ping: hostname, OS, RAM, CPU — into our operator chat. It does not need admin, does not open a port, and does not download the model by itself. Piping a URL into bash means you trust this site. Safer: download the script, read it, then run it.

What it is not: a sandbox for your private files. Don’t host on the laptop that holds taxes and iCloud if that makes you uneasy. Use a leftover Air that can stay on a desk, plugged in. Heat and power are the real cost, not “someone hacking the Air through a prompt.”

How do I get paid for hosting a model?

In digital dollars, from two places. First: a regular payment from a cut of coin trades, if your model is loaded, your computer is online, and our spot-checks still pass. Second: extra pay when your computer answers a request — about 70% of what that request costs. Subscriber money does not go to you; we use it to buy the coin. A computer that is only turned on, with no model, is not paid.

How much will I earn?

It depends on which model you run, how often you stay online, how much the coin is being traded, and whether anyone is sending requests. Do not quit a job over this.

How do I get paid?

Earnings accrue in USDC and pay out automatically once they cross a small threshold, through a regulated payment provider. Expect a one-time identity check at your first payout — that's the provider's compliance requirement, not ours.

What does it cost me?

Electricity while a model is loaded and while jobs run. You choose which model to host and you can stop any time. In-flight work moves to another device.

Buying

Who buys this? Is it like subscribing to ChatGPT or Claude?

ChatGPT and Claude are products for a human in a window. Our customer is often an agent — software that already speaks chat-completions, can pay, and will hit us all day. It subscribes (that USDC buys the token) and it sends jobs (that USDC pays the host who served them). Humans can subscribe the same way. Same pool.

Why is it cheaper than the big providers?

Their price covers datacenters they had to build and power. Ours covers a fraction of the electricity on computers that were already bought and already on. Same free models, structurally cheaper hardware. That's the entire edge, and it doesn't expire.

How do I know the answers are honest?

A slice of all jobs is quietly re-run on a second, independent device and compared. Devices that match build trust and get more work; devices that don't stop earning. Hosting stipends also require probes to keep passing. Verification on real, mixed hardware is still being proven — that is phase 5, not a finished claim.

How do I integrate?

If you have a chat-completions client, an agent harness, goose, Codex, whatever — change baseURL to us. That’s the whole integration. Streaming included. The agent doesn’t need a special Reasoning Capture SDK.

Should I send sensitive data?

Treat us like any inference provider: the device serving your request sees its text. Don't send secrets you wouldn't send an AI API generally. Encrypted execution is on the roadmap; today the honest answer is: keep regulated data off the network.

What about latency-critical or frontier-quality work?

Not our fight. If you need sub-second responses in a user-facing product or frontier closed-model quality, use a datacenter provider. We win on price for work that can wait seconds, not milliseconds.

The money layer

Why USDC and not points?

Points counted on your own device can be faked, and points aren't money. USDC is a dollar. All metering happens on our servers — a device can't inflate its own earnings, and neither can we quietly dilute yours.

What's the token?

A coin on Solana, launched in public on pump.fun. It is not your paycheck. Subscription money is used to buy it. Trade fees pay hosts in digital dollars. You do not need to own the coin to earn or to subscribe. This site is not selling it and is not investment advice.

Does the network pre-buy AI credits from anyone?

No. The core loop has no upstream supplier: buyer pays, a device serves, the owner earns, the network keeps a share. During the bootstrap period some frontier-model requests route through a cloud provider at a margin — a convenience tier that shrinks as network devices come online, never a subsidy.