One endpoint you already know, plus the platform around it

Authentication

Every request carries your Shaidle token. Both header forms are accepted, so a client written for either provider convention works without modification:

Authorization: Bearer sk-shaidle-…
# or
x-api-key: sk-shaidle-…

Tokens are minted in the app under Connect editor, shown once, and can be revoked individually. Revoking a token ends that client's access immediately; it has no effect on any provider account behind the exchange.

Wire-compatible endpoints

Example

curl https://<your-exchange-endpoint>/v1/messages \
  -H "Authorization: Bearer sk-shaidle-…" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Explain this stack trace."}]
  }'

model takes any model the exchange lists — Claude and Codex models are priced from the same rate card, and the one you name decides which lane serves you. The current card is on the pricing page.

Streaming works exactly as it does against the provider. If the seat serving you runs out of quota mid-request, the exchange fails over to another seat before any byte has been sent to you — you see one clean response, not a truncated one.

The Codex side

curl https://<your-exchange-endpoint>/v1/responses   -H "Authorization: Bearer sk-shaidle-…"   -H "content-type: application/json"   -d '{
    "model": "gpt-5.6-codex",
    "input": "Explain this stack trace."
  }'

The same exchange, on the wire OpenAI-side tools speak. Point the Codex CLI at it with a named provider in ~/.codex/config.toml (wire_api = "responses") — the Connect your editor screen prints the exact file. You stay signed out of ChatGPT; every turn is served from a connected seat and billed to your wallet.

Ask GET /v1/models with your token for the models the exchange can actually serve — it is the union of what connected seats report, not a list we maintain, so it follows the providers rather than lagging them.

Platform API — buyers

Platform API — sellers

Seat reporting is opt-in and sends counts, never content. The reasoning is in the compliance position.

Errors

Each wire endpoint answers in ITS OWN provider's error shape — Anthropic's {"type":"error","error":{…}} on /v1/messages, OpenAI's {"error":{"message":…,"type":…}} on /v1/responses — so an unmodified client renders the message natively instead of showing a parse failure.