Wokey

What it is: a third-party LLM API aggregation and relay gateway. It does not train or own any model; requests are ultimately handled by the upstream vendors' official endpoints. Chinese communities often call it "the Pinduoduo version of OpenRouter".

One key, one base_url change

It speaks three protocol shapes: OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. Migration usually means swapping the base URL, the API key and the model ID — note that IDs are hyphenated, such as claude-sonnet-5-5, not OpenRouter's anthropic/claude-sonnet-5.5. Official setup guides exist for Claude Code, Codex CLI, Claude Desktop, opencode, Grok CLI and OpenClaw, and the gateway accepts request paths both with /v1 and without /v1.

The differentiator is verifiability

Most relays can only promise in words that you are getting the official model. Wokey ships evidence you can check offline instead: the official exit runs inside an AWS Nitro Enclave, pass-through responses carry an Ed25519 signature plus attestation, the signing key is bound to the Nitro root through a reproducible image measurement (PCR0), and the verifier is open source under Apache-2.0, so you can confirm on your own machine that the response bytes were not altered and that the upstream host really is the vendor's domain.

The boundaries matter: only pass-through routes carry proof. Requests that go through format conversion (calling Claude over Chat Completions, for example), image generation, and responses served while the enclave is degraded carry none. The request side is unprotected too — truncating history or lowering max_tokens will not be caught by a response proof. The vendor's own rule of thumb: treat any response without tee.proof as unproven.

Pricing structure

Pure pay-as-you-go against a prepaid balance, with no subscription and no monthly fee. Every Claude billing line is priced at 15% of the official reference, so the discount holds regardless of cache hit rate; flagship GPT and Grok models land around 85% off; open-weight and Chinese models are discounted far less steeply, and a few have essentially no discount. Some DeepSeek models are billed at peak and off-peak rates, and multi-chain USDT top-ups carry no fee.

Who it fits

Heavy individual users of Claude Code or Codex, small teams that need to run batch or long-context jobs cheaply, and anyone who wants to validate the service with a small balance before migrating. It does not fit enterprise procurement, buyers who need a DPA, SLA or SSO, or any workload whose data cannot leave its jurisdiction — there is no enterprise contract, and a relay necessarily sees plaintext.

Editorial note and risks

What is listed here is the third-party aggregation layer itself, not a claim that its upstream relationships are vendor-approved. Its low prices come from sharing idle consumer subscription quota: on the provider side, users authorise consumer subscriptions over OAuth and the platform dispatches requests to them. That sits in direct conflict with upstream terms that forbid sharing credentials, reselling access and intermediating usage, so supply can be cut off at any time. Top-ups are non-refundable, the operating entity and jurisdiction are undisclosed, and liability is capped at the greater of fees paid and 100 USD. Traffic passes through in plaintext, and the privacy policy carves out an exception where abnormal requests may keep the full request body for up to 14 days. Practical advice: top up only what you can afford to lose, split keys per client so you can revoke one independently, and always keep an official direct route as a fallback.