Client / Router / Network
Noema is building a private AI routing layer: a local-first Client with ZKR today, growing into attested GPU compute and an onion-routed, decentralized Network.
Noema is private AI routing infrastructure, delivered in stages through the Client, Router and Network. The Client is the workspace for writing, reasoning and code, with control over where a request runs. The wider vision connects local intelligence to confidential compute rather than requiring every task to go to a conventional cloud model.
Run an open model in Your Browser with WebGPU, connect Ollama or a Local API on Your Machine, or choose Noema Router for automatic cloud access when available. Noema Router requires no user API key or cloud model selection. Chats, attachments and optional memory are stored in your browser. Reply receipts report where processing happened. Web search is optional and sends a query out.
The ZK Redaction Engine detects supported sensitive text, replaces it locally before a supported outgoing request, then restores matching original values in your reply. ZKR is best-effort redaction, not a cryptographic zero-knowledge proof. PDFs, text-only DOCX documents and English receipt images can be read locally and reviewed before sending protected text and rebuilt pages. Original pixels, codes and metadata are excluded. Some formats need a clearer export. Originals and local memory are not encrypted merely because they are stored locally.
Noema Router brings automatic cloud selection and fallback into the Client under one route. It is labelled Cloud because those requests leave the device. It is a step toward the routing product, separate from the local Engine and the future confidential Network. It does not mean Noema trained the underlying models or that cloud requests use an enclave.
The first Network service is planned around a Noema-operated confidential GPU, protected memory, remote attestation checked by the client, and an independent relay. Attestation provides evidence about the measured hardware and software. Software inside a TEE still processes plaintext; attestation alone does not prove that software cannot log or that its answers are correct.
The first Network service uses a single relay to separate client addressing from the compute endpoint. The wider Network adds three-hop onion routing and independent operators. Layered encryption is intended to keep any single non-colluding relay from seeing both the client address and the final workload. A single relay is not the completed onion network, and the current cloud Router does not provide onion routing.
Following document, image and receipt ZKR, the sequence is in-app generated-tool previews and useful workflows; the first attested Network service; then the wider operator network. The wider phase includes three-hop transport, operator staking and enforceable slashing. Multi-model verification is a later optional check for disagreement and evidence, not a guarantee of truth. The Provider Node and SDK follow usable Network interfaces.
NOEMA is live on Robinhood Chain. Inference settlement, operator rewards, staking and slashing are planned Network utility. A live token does not mean those services are operating. The official contract is 0x4f9be3c04d1de309a83d7186f537d4c012d52d41.