What is IO?

IO is the first AI inference protocol that cryptographically proves it retains nothing. Every response carries an Ed25519-signed receipt. You don't trust us — you verify the signature.

Input. Output. Nothing.

Built for privacy-first AI — uncensored, ephemeral, and structurally incapable of storing your data.

The Problem

Every conversation with an AI today leaves a record you don't control. Your prompt, your phrasing, the half-formed idea, the secret pasted in haste, your IP address — logged, linked to an identity, retained on a schedule you can't see, and often folded back into the next model's training set.

The most intimate interface we've ever built runs on a business model of remembering everything. Privacy, where it exists, is a line in a terms-of-service document — not a property of the system. You can't inspect a provider's storage layer. You can't falsify a "we deleted it" claim.

IO inverts the premise. The surest way to protect a thought is to run on a system that was never able to keep it. No database sits in the request path; inference happens in memory and is wiped on response. And because the absence of a thing is hard to prove, every answer ships with a signed receipt that commits to the conditions it was produced under.

You don't trust the policy. You verify the signature.

Privacy alone isn't a product, though. So IO pairs verifiable non-retention with a real reason to switch: a multi-model orchestration engine that broadcasts one prompt across the best open-weight models, synthesizes their answers with attribution, and lets you branch reasoning the way you branch code. Verifiable privacy and model diversity — as a single stack.

Core Guarantees

• Zero retention architecture — no databases, no logs, no analytics, no IP retention, no accounts. Not a policy. A constraint.
• Cryptographic receipts — every inference returns a signed io_receipt_v1 proving retained_prompt: false. Verify against the published public key at useio.xyz/receipt.pub.
• Self-destruct at every level — session, message, file, canvas, room. Timers from 1 minute to flash (tab close). No recovery path. Gone means gone.
• Uncensored by design — open-weight models on a permissionless inference layer. No guardrails, no filters, no policy engine.
• No accounts. No KYC. — drop an API key or use a relay bot. Zero friction.

Architecture

Every request flows through the same path: redact, route, infer, sign, wipe. The model-orchestration side of this pipeline — broadcast, diffuse, fork/merge — is covered in depth in AI Architecture.

What You Can Build

Feature Matrix

Why "IO" vs Everything Else

Most "private AI" products ask you to trust their word. IO makes trust unnecessary — every response is signed, every claim is verifiable, every conversation is structurally ephemeral.

• DOT has ~8 MCP tools → IO has 21
• DOT has no broadcast, diffuse, diff, batch, fork/merge, ReceiptChain, redact, proof certs, warrant canary, or SSE streaming
• DOT supports 1 SDK language → IO ships 3 (Rust, Go, TypeScript)
• DOT has no relay → IO covers 4 platforms (Telegram, WhatsApp, Signal, Discord)

Token & Payment

Every inference is metered and settled through the x402 micropayment protocol. The network token, $IO, is the unit of access and payment — fair launch on PumpFun (Solana), no pre-sale, no VC allocation.

• Pay per call — completion, broadcast, proof certificate, redact scan. Pay in $IO for a discount.
• Access gate — broadcast/diffuse, higher rate limits, and priority routing require $IO.
• Buyback-and-burn — a portion of net protocol revenue (after GPU + bandwidth costs) is used for $IO buyback-and-burn.
• Governance at v1.0 — $IO holders vote on the active model registry, redaction policy, fee splits, and the warrant-canary cadence.

Fair launch. No insiders. No unlocks. The token demand is a function of real inference usage, not speculation alone.

Next Steps

Ready? Head to Quick Start to run your first inference.