Three lines to proof.
Call any registered model. The proof returns with the result.
One client
Pass a registered model's name and your input; the answer and its proof come back together. There is no separate proof endpoint.
TypeScript
The SDK for Node, Deno and the browser.
npm install @aios/sdkPython
The same client for Python services and notebooks.
pip install aios-sdkREST
No SDK needed: the same call and the same proof over HTTPS.
POST /v1/inferimport { AIOS } from "@aios/sdk"
const aios = new AIOS(key) // scoped, capped, revocable
const r = await aios.call("bitnet-ref-1", prompt)
r.output // the answer
r.proof // job id, model, mode, committee, settlement
Every field
Provenance of execution
The record proves this model produced this output, unforged. On large models the assurance is statistical, and the record says so outright.
- proof.job the job id, to look up on the explorer
- proof.model the registered name and address: no silent downgrade
- proof.mode verified or attested, honest by schema
- proof.committee size, matches and the canonical hash
- 01 / Key
Mint a scoped key
In AIOS:PLATFORM: a bounded-spend credential, capped and revocable, never custody.
- 02 / Resolve
Name a model
The SDK resolves address, price and mode from the registry. No hard-coded endpoints.
- 03 / Call
Submit and settle
The call escrows in $AIOS and clears by committee consensus.
- 04 / Keep
Keep the proof
The record is part of the return value. No quorum: a typed error and a full refund.
The toolchain
Docs
The API reference, the verification record schema and every error type.
AIOS:PLATFORM
The console and REST API, where keys are minted and corpora published.
AIOS:MCP
The agent doorway, a verification record on every call.
Registered is not endorsed
A model you can call is not a model anyone vetted. The mode says what proof a call earns, nothing more.
Answers you can check
Call a registered model, read the proof, and look the job up on the explorer yourself. That is the whole promise.