Introduction

Apex Ant is a Venice inference router sold on the AntSeed peer-to-peer network. You get frontier models (text, image generation and editing) at a hard-capped share of Venice list price, with pay-as-you-go USDC settlement on Base. No accounts in the traditional sense: a wallet signature is your identity.

Your tool (chat UI, curl, agent)
  ↓
Apex Ant gateway / your own AntSeed buyer
  ↓ encrypted P2P
Apex Ant seller node
  ↓
Venice zero-retention inference

What makes it different. Prices are public and provable: the board shows every model against its Venice reference, refreshed continuously. Funds sit in an on-chain escrow contract, not on our books. And the exit is always open: your buyer key is exportable, the protocol is open.

Quickstart

Three minutes, no installs:

1. Open the cabinet, connect any EVM wallet (MetaMask, Rabby, WalletConnect).
2. We generate a dedicated buyer identity for you. Fund it with USDC on Base via the deposit button.
3. Create an apx_… key. Paste it into the chat, or point any OpenAI-compatible tool at our base URL.

curl https://apex-ant.net/api/v1/chat/completions \
  -H "Authorization: Bearer apx_…" \
  -H "content-type: application/json" \
  -d '{"model":"gpt-6-astra","messages":[{"role":"user","content":"hi"}]}'

Two paths

Same seller, same models, same prices, same settlement. Choose per threat model; you can use both at once.

Our honest recommendation: hosted is the easier start: browser, one click, familiar API key, and your buyer key sits inside a hardware-attested TEE enclave running publicly auditable code (not just "trust our server"). Native is still the more robust endgame: fully trustless, your keys never leave your machine, no dependency on our infrastructure at all. Many users start hosted and move to native once the workflow proves itself; your buyer key exports either way.

Hosted path: cabinet

Sign-in

SIWE-style: sign a challenge with your wallet, receive a session token. No email, no password. The wallet is only an identity; it never pays for requests directly.

Buyer identity

First sign-in generates a dedicated AntSeed buyer (secp256k1 keypair) bound to your wallet. It opens payment channels and pays for requests. One wallet → one buyer.

Deposits

USDC on Base into the AntSeed deposits contract, credited to your buyer. The cabinet button handles chain switch → approve → deposit. Direct USDC transfers to your buyer address are auto-swept into your deposit within ~5 minutes.

API keys

Up to five per wallet, stored as bcrypt hashes, shown once. A key spends only from your own buyer's deposit.

Backup & exit

Export the buyer's private key from the cabinet at any time. With it you can withdraw or operate the buyer without us. We cannot recover a lost key, and cannot block an exit.

Roadmap note: the buyer stack already runs in an attested TEE enclave (see privacy); moving the seller-side inference stack into a TEE as well is planned.

Native path: AntSeed P2P recommended

1. Install the AntSeed VPR desktop client (recommended) or the CLI node for headless setups.
2. Fund your buyer with USDC on Base.
3. Pin our seller: 73b4c933…4de736 (click to copy the full peer id).
4. Your client discovers our live catalog and prices from the network. Nothing else to configure.

Your keys never leave your machine. Auth is per-request signatures; settlement is the same on-chain escrow both paths share. This path keeps working even if our entire hosted stack disappears. The network, the escrow and your funds don't care.

API reference

Base URL https://apex-ant.net/api/v1 · auth Authorization: Bearer apx_…

Web search

Venice-style parameters pass through: {"venice_parameters":{"enable_web_search":"auto"}}. Citations arrive as search_citations.

Errors

OpenAI error shape. 401 bad key · 402 balance empty · 404 model_not_found · 429 rate_limited (honor retry-after) · 503 upstream failure, retry.

Web chat

The bundled chat is our build of the open-source Venice frontend: streaming, reasoning view, web search with citations, image generation and editing, multi-model playground. Your apx_… key lives in browser local storage and goes only to our API. The header shows live balance (available / in-sessions), refreshed every 15 seconds; model pickers show prices inline (in / out / cache per 1M tokens).

Images & edits

Generate: ~47 models (gpt-image-2, flux-2-max, nano-banana-pro…) billed per image. Edit: ten *-edit models (qwen, flux, gpt-image, nano-banana, seedream): base64 image + prompt in, edited image out, billed per output image at 40% of the Venice reference. Upscale and background removal ship when the settlement protocol supports them.

Billing & balances

• Pay-as-you-go. Deposit USDC once; each request is metered and covered by off-chain signed authorizations against your deposit.
• On-chain escrow. Funds live in the AntSeed deposits contract. Settlements are real Base transactions.
• In-sessions. The reserve locked by an open channel. Releases on idle (~10 min) or instantly via «Close session» in the cabinet.
• Channels. Cabinet lists every channel: authorized, settled, status. Reconcile with any Base explorer.

On-chain

Base mainnet (8453), USDC:

Public settlement: seller page on antseedstats.

Pricing model

Sell price tracks the going rate of the cheapest Venice-inference lanes plus our fixed markup, hard-capped at 50% of Venice list, typically much lower. The price board is the promise: every row shows our price, the Venice reference and the live ratio, continuous refreshes, 7-day history. No free tier: every token is paid, which is exactly what keeps the economics honest.

Privacy

• ZDR upstream: inference runs on Venice's zero-data-retention fleet; prompts are processed in memory, never written or trained on.
• We don't log content: the edge stores operational metadata only (model, tokens, latency), never prompt or completion bodies.
• No account PII: a wallet signature is the whole account. No email, no trackers, no analytics beacons.

Full statement: privacy.

Key custody, honestly

Hosted needs a hot key to sign your payment authorizations: a copy of your buyer key lives on our server, AES-256-GCM encrypted, key material in environment only. That is the honest trade of one-click UX. Mitigations: your exposure is bounded by your deposit; the key can never touch your wallet; export-and-exit works at any time. The hosted stack runs inside a TEE (Phala CVM) with live attestation on the privacy page: even we can't read the keys. If the trade doesn't fit you today, the native path is trustless by construction.

Limits & errors

Rate-limited responses carry retry-after. Streams hold a slot until closed.