Aile is a marketplace for AI capacity. Sharers list capacity they already have and set their margin. Callers reach that capacity through compatible APIs and pay per call in USDC. There is no Aile subscription.
This note follows one request from the moment a caller sends it to the moment the sharer who served it gets paid.
Three ways to share capacity
Self-hosted. The sharer runs a model on their own hardware, through Ollama, llama.cpp or another compatible server. The weights stay on that machine. The sharer's machine reads the prompts it answers, because answering them is the work.
Through a node. The sharer links a provider subscription or API key. Aile stores the credential encrypted at rest and injects it when serving. The sharer's node forwards encrypted traffic and can't read the prompt or the credential. Aile's relay can read the request and the response.
Nodeless. Aile connects to the provider directly, with no sharer node in the path. Nothing needs to run on the sharer's computer. That also means switching the computer off doesn't stop serving. The sharer has to turn nodeless off. API-key accounts can go nodeless from the CLI, and the dashboard can also serve a subscription nodeless. That second case carries more risk for the sharer, because it is still one person's plan.
Linking an account gives Aile technical access to it, not permission to resell it. Many providers' terms restrict sharing a credential, and sharers must be allowed to share what they list.
From request to response
A caller sends /v1/chat/completions from an OpenAI-compatible client or /v1/messages from an Anthropic-compatible one. The path sets the message format, and the model id sets the provider. cc/claude-sonnet-5 goes to Claude Code capacity and nowhere else. If no sharer for that provider is online, the caller gets a 503 instead of a quiet switch to another provider.
Before anything is charged, Aile swaps recognised secrets in the prompt, such as API keys and private keys, for placeholders and runs a narrow content check. A refusal at this stage costs nothing.
Then the router builds the pool. It drops sharers that don't fit the request's model, price ceiling or other limits. Among those left, price comes first: the cheapest sharer is tried first. At the same price, tested sharers go first, then a draw weighted by recent health and speed. Aile dials up to four sharers per request. Once a response starts streaming, it stays with the sharer who started it. A conversation also returns to the sharer who served its last turn.
Callers have several limits they can set:
max_tokenscaps the output.x-aile-max-pricecaps the rate.x-aile-max-costcaps a single request in dollars.- A spend limit on each API key resets daily, weekly or monthly.
How payment works
Balance. This path works with any ordinary API client. Aile reserves an estimate before the call, commits the real usage afterwards and releases the rest. Sharer earnings settle in batches. Aile's fee is 10% on node-served calls and 15% on nodeless calls, where 5 points cover egress. The fee comes out of the sharer's share. It is never added on top of the caller's price.
x402. A wallet-enabled client signs a USDC payment for one request, with no Aile account needed. It reaches self-hosted models and sharers selling through an API key, not subscriptions. The payment settles after the answer is ready and before it's sent, so a failed answer isn't charged. 100% goes to the sharer.
A signed payment names one sharer at one price. If that sharer fails, the payment doesn't move to someone else. Answers on this path are buffered, not streamed.
What sharers control
A sharer chooses what to offer and sets a margin of each model's published list price. The margin runs from 0 (free) to 1 (list), overall or per model. A sharer can turn any model off and turn it back on. Pricing yourself too high doesn't cause an error. Requests just go to someone cheaper.
Earnings aren't the same as profit. A sharer with an API key still pays their provider's usage charges, and a self-hosted sharer pays for power and hardware. A margin has to cover those costs and Aile's fee.
What Aile can and can't verify
Every chat reply names the sharer (x-aile-sharer), the capacity type (x-aile-capacity) and the price. x-aile-blind says whether the sharer's node could read the request. It doesn't mean Aile's relay couldn't.
Aile can check that a credential recently worked. For some providers it can also prove the account link was real. Listed models must pass a test request, and they're re-tested about once a day. Self-hosted listings get a known-answer probe.
None of this proves which model produced a given answer, or how good it was. We don't have broader evaluations or a reputation system yet.
So start small: one request, a spend limit on the key, and input you wouldn't mind anyone reading. Check both the answer and its recorded cost before you send more.