Services · inference-gateway

PPQ.AI (PayPerQ)

Pay-per-query access to 500+ frontier models — image, video and data calls need no account at all (pay the Lightning invoice in the 402 and go); chat and audio need a key and a funded balance, which Nostr Wallet Connect can top up on its own.

What it is

PPQ.AI (PayPerQ) is a gateway to 500+ frontier chat, image, video, and audio models (OpenAI, Anthropic, Google, xAI, Mistral, Meta, Deepseek, Qwen, and more) behind one OpenAI-compatible API, paid per use with no subscription. Average query cost is around 1.5¢, with top-ups from about 10¢. It is a live instance of “an agent pays for its own inference” — the simplest on-ramp for an agent to buy frontier-model calls without provisioning an account with every provider separately.

What makes it worth the account anyway. The balance is funded over Lightning, on-chain Bitcoin, Liquid BTC, Litecoin, Monero, USDT/USDC, or a card — and, the part that matters most for an autonomous agent, it can be topped up by Nostr Wallet Connect: connect an NWC wallet (Alby Hub, LNbits, Primal) and PPQ refills the balance itself when it drops below a threshold. That is an agent buying its own inference indefinitely without a human returning to the payment screen — reached through an account rather than around one. The Bitcoin, Lightning and Liquid paths settle in Bitcoin; the stablecoin and card paths settle in assets an issuer can freeze, and the card path is a KYC surface.

It also runs a private-inference tier: a set of models served inside TEEs (trusted execution environments, on NVIDIA confidential-computing GPUs) with end-to-end encryption — the request body is encrypted in your own process (by the local proxy, using HPKE) before it reaches PPQ’s servers, decrypted only inside the hardware-attested enclave, and encrypted again on the way back, so PPQ “can only see ciphertext and routing metadata — never prompts or completions.” Private models are addressed with a private/ prefix and currently include private/glm-5-2 (384K context), Kimi K2.6, GPT-OSS 120B, Llama 3.3 70B, Qwen3-VL 30B, and Gemma 4 31B. It also publishes a /llms.txt and an API surface designed to be consumed directly by agents.

When to use it

• Accountless image, video or data-enrichment calls — pay the 402’s invoice and go, nothing to sign up for.
• Agents that need to buy inference on demand without managing a separate account with every model provider — one key here instead of eight elsewhere.
• Unattended agents that must not run out of credit — NWC auto-topup refills the balance from a Lightning wallet without a human.
• Privacy-sensitive workloads — the private/ TEE tier keeps prompts unreadable to the gateway.
• Lightning-native metering of model usage.

Dependencies & payment

Dependencies — and they differ by endpoint. For images, video and data enrichment: a Lightning / L402 wallet or L402-capable client, and nothing else — no account, no key. For chat and audio: an account, an sk- API key, and a funded balance. For private/ models: all of the above plus the local encryption proxy, which is required rather than optional (Node 18+; it handles enclave attestation and the HPKE key exchange, powered by Tinfoil). Payment: pay-per-query, ~1.5¢ a query, top-ups from ~10¢ — funded over Lightning, on-chain Bitcoin, Liquid BTC, Litecoin, Monero, USDT/USDC, or card, with NWC auto-topup for unattended refills. Card funding adds KYC; the crypto paths do not.

Quick start

Which path you take depends on what you are buying.

Images, video, data enrichment — no account. Send the request with no Authorization header, pay the Lightning invoice that comes back in the 402 challenge, and replay the request with Authorization: L402 <token>:<preimage>. Clients like lnget and mppx automate the challenge-pay-replay loop for you.

Chat and audio — account, key, balance. Sign up at ppq.ai, generate an sk- key at ppq.ai/api-docs, fund the balance (Lightning is the fastest, and NWC auto-topup keeps it funded), then call it like OpenAI:

curl https://api.ppq.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-YOUR_API_KEY" \
  -d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"Hello"}]}'

Chat, image, video and embeddings sit at https://api.ppq.ai; audio is on a different base URL, https://ppq.ai/api/v1.

Private inference — the proxy is required, not optional. You cannot call a private/* model against api.ppq.ai directly; PPQ says so in as many words. Run the open-source proxy locally and point an OpenAI- or Anthropic-SDK client at it — it handles attestation, the HPKE key exchange and the encryption:

PPQ_API_KEY=sk-YOUR_API_KEY npx ppq-private-mode   # starts on 127.0.0.1:8787

The proxy also speaks the Anthropic Messages API, so Claude Code and other Anthropic-SDK clients work against it directly. The core service itself is not open-source. Machine pointer: ppq.ai/llms.txt.

Gotchas

• The accountless L402 path does not include chat. See the callout above — this is the single most important thing to check before designing around this service. Image, video and data-enrichment endpoints are accountless; LLM calls are not.
• It is a centralized custodial gateway to third-party providers — PPQ proxies your prompts upstream. Only the private/ TEE tier (hardware-attested enclave + client-side encryption) keeps content unreadable to PPQ; the standard models do not.
• The private tier is not more sovereign, it is more private. It still needs an sk- key and a funded balance, and it adds a local proxy process you have to keep running. It reduces what PPQ can read, not what PPQ can withhold.
• The no-account L402 path stores the payment credential on your side — lose the token/preimage and you lose the call you paid for.
• Funding by card introduces KYC; the crypto funding paths do not. NWC auto-topup is the one to reach for if the agent runs unattended — otherwise a drained balance stops it silently.