One URL for every tool

The wire, precisely

Every call is POST /api/v1/chat/completions with header Authorization: Bearer hk-key and a JSON body of model plus messages. The gateway answers the OpenAI shape: choices with message content, plus usage with prompt_tokens and completion_tokens. Anything that parses an OpenAI response already parses ours — no adapter, no SDK fork.

Model ids are namespaced hooder/. Examples live today: hooder/chatgpt-terra for volume work, hooder/chatgpt-astra for flagship GPT-class calls, hooder/claude-fable-5 and hooder/claude-fable-5.1 for Claude-class calls, hooder/claude-opus-5 for deep reasoning. The full callable list is always GET /api/v1/models — the docs never drift from it because the page reads the same table.

POST /api/v1/chat/completions
Authorization: Bearer hk-key
{"model":"hooder/chatgpt-terra","messages":[{"role":"user","content":"halo"}]}
# 200 + choices[0].message.content + usage{prompt_tokens, completion_tokens}

Keys are the access control

A key is not a password — it is a scoped credential. Each key carries four controls: an allowlist of model ids, a weekly token budget, a traffic class, and an expiry timestamp. A call outside the allowlist is rejected before it touches upstream; a call over budget fails fast with a clear error instead of silently spending.

Revoke is immediate. The bearer hash lookup happens per request, so revoking a key in the console kills it on the very next call — no propagation delay, no session to drain. Minting a replacement takes seconds and does not disturb sibling keys.

Per-tool notes

OpenCode / Cline / Continue / Roo Code: set the provider base URL to https://hooderai.site/v1 and paste the hk-key as the API key. Pick a hooder/ id from /v1/models. Nothing else changes — prompts, streaming flags, and temperature pass through untouched.

Codex CLI / Claude Code / OpenClaw: same base URL, same key. These harnesses tend to burn tokens fast on agentic loops, so put them on a dedicated key with a weekly budget — when the loop runs away, the budget stops it, not your balance.

Zed: configure the OpenAI-compatible provider with the Hooder base URL. Zed's inline-assist calls are small and frequent, which is exactly the shape metered billing favors over subscriptions.

Failure modes, honestly

401 means the key is wrong, revoked, or expired — check the prefix in the console. 403 means the model is outside this key's allowlist. 404 means the model id does not exist or is disabled — re-pull /v1/models. 502 means inferhub upstream was unreachable; the attempt is logged with error status and zero price.

Failed calls are never billed. They are counted in usage with status error and price $0, so your debugging does not cost you and your invoice stays reproducible.