What you send

The public contract is OpenAI-shaped JSON. POST /v1/chat/completions works for every alias on the models page, including Claude and Gemini. POST /v1/responses (and the same path as a WebSocket) is for OpenAI models and grok-4.6 — those two providers already speak Responses natively. Claude aliases also answer natively on POST /v1/messages (Anthropic Messages shape); Gemini stays translated at the edge, and there is no public /v1/interactions. Translated calls still read OpenAI cached_tokens for billing.

Chat Completions

Point any OpenAI client at the gateway. Same shape for GPT-5.6, grok-4.6, Claude, and Gemini.

ts

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://gateway.codegraff.com/v1",
  apiKey: process.env.CODEGRAFF_API_KEY, // cg_sk_...
});

const res = await client.chat.completions.create({
  model: "grok-4.6", // or gpt-5.6, gemini-3.7-flash
  messages: [{ role: "user", content: "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}" }],
  reasoning_effort: "high",
});
console.log(res.choices[0].message.content);

bash

curl https://gateway.codegraff.com/v1/chat/completions \
  -H "Authorization: Bearer cg_sk_your_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6","messages":[{"role":"user","content":"Hello"}]}'

Responses (OpenAI + Grok)

Official Grok and GPT examples use client.responses.create. That maps 1:1 on the gateway for gpt-5.6 / gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna and grok-4.6. We rewrite the alias to the upstream model id and POST to api.openai.com/v1/responses or api.x.ai/v1/responses. Claude returns 400 here — use /v1/messages for that. Gemini stays on Chat Completions.

ts

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://gateway.codegraff.com/v1",
  apiKey: process.env.CODEGRAFF_API_KEY,
});

const response = await client.responses.create({
  model: "grok-4.6",
  prompt_cache_key: "session:bugfix-v1",
  input: [
    {
      role: "user",
      content:
        "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}",
    },
  ],
});
console.log(response.output_text);

bash

curl https://gateway.codegraff.com/v1/responses \
  -H "Authorization: Bearer cg_sk_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6",
    "prompt_cache_key": "session:bugfix-v1",
    "input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
  }'

Messages (Claude)

POST /v1/messages speaks Anthropic's Messages API end to end for the Claude aliases (claude-opus-5-5, claude-sonnet-5-5) — no translation. Thinking blocks and their signatures replay across turns, thinking.display summarizes as upstream intends, images inside tool results work, and cache_control — including "ttl": "1h" — is honored.

Auth is x-api-key: cg_sk_... (the Anthropic SDK default) or Authorization: Bearer cg_sk_.... max_tokens is required — 0 works as a cache pre-warm — and there is no gateway cap, the model's own maximum applies. Billing: input tokens at list, cache reads at the cached rate, 5-minute cache writes at 1.25× input, 1-hour writes at 2× input. Anthropic server tools (web search, code execution, computer use, and the rest, including mid-conversation tool additions) are rejected, as are mcp_servers and container. The forwarded anthropic-beta set is thinking-binding-controls-2026-08-01, thinking-display-updates-2026-08-18, mid-conversation-output-config-2026-07-01, mid-conversation-tool-changes-2026-07-01, mid-conversation-system-clear-at-2026-08-21, inline-tools-2026-09-15, and context-management-2025-06-27. Fast mode ("speed": "fast") is not offered yet — it returns a 400. A non-global inference_geo is rejected; service_tier passes through. Request errors come back with Anthropic's own error type and message. Non-Claude aliases return 400 here — use Chat Completions.

POST /v1/messages/count_tokens is free and takes the same body minus the billing checks.

ts

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://gateway.codegraff.com",
  apiKey: process.env.CODEGRAFF_API_KEY, // cg_sk_...
});

const res = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 64000,
  thinking: { type: "adaptive", display: "summarized" },
  messages: [{ role: "user", content: "Find and fix the bug, then explain it" }],
});
console.log(res.content);

bash

curl https://gateway.codegraff.com/v1/messages \
  -H "x-api-key: cg_sk_your_key" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 64000,
    "thinking": {"type": "adaptive", "display": "summarized"},
    "messages": [{"role": "user", "content": "Hello"}]
  }'

WebSocket

GET /v1/responses with Upgrade: websocket keeps one client socket open. We accept that immediately, then open an upstream WebSocket to OpenAI or xAI on the first response.create — after we know which host to use. If the upstream WS fails, that turn falls back to HTTP/SSE; your socket stays up. Switching from gpt-5.6 to grok-4.6 on the same client socket reconnects upstream.

Auth: Authorization: Bearer cg_sk_... from Node, or pass the key as a WebSocket subprotocol from a browser (browsers cannot set that header). Each block below has a copy button.

ts

// gpt-5.6 — Node 22+
const ws = new WebSocket("wss://gateway.codegraff.com/v1/responses", {
  headers: { Authorization: `Bearer ${process.env.CODEGRAFF_API_KEY}` },
});

ws.addEventListener("open", () => {
  ws.send(JSON.stringify({
    type: "response.create",
    model: "gpt-5.6",
    prompt_cache_key: "session:ws-gpt56",
    input: "Reply with the single word pong",
  }));
});

ws.addEventListener("message", (ev) => {
  const event = JSON.parse(String(ev.data));
  if (event.type === "response.output_text.delta") process.stdout.write(event.delta);
  if (event.type === "response.completed") ws.close();
});

ts

// grok-4.6 — Node 22+
const ws = new WebSocket("wss://gateway.codegraff.com/v1/responses", {
  headers: { Authorization: `Bearer ${process.env.CODEGRAFF_API_KEY}` },
});

ws.addEventListener("open", () => {
  ws.send(JSON.stringify({
    type: "response.create",
    model: "grok-4.6",
    prompt_cache_key: "session:ws-grok46",
    input: "Reply with the single word pong",
  }));
});

ws.addEventListener("message", (ev) => {
  const event = JSON.parse(String(ev.data));
  if (event.type === "response.output_text.delta") process.stdout.write(event.delta);
  if (event.type === "response.completed") ws.close();
});

ts

// Browser — key must be a subprotocol, not an Authorization header
const KEY = "cg_sk_your_key";
const ws = new WebSocket("wss://gateway.codegraff.com/v1/responses", [KEY]);

ws.onopen = () => {
  ws.send(JSON.stringify({
    type: "response.create",
    model: "gpt-5.6", // or "grok-4.6"
    input: "Reply with the single word pong",
  }));
};

ws.onmessage = (ev) => {
  const event = JSON.parse(String(ev.data));
  if (event.type === "response.output_text.delta") {
    document.body.append(event.delta);
  }
};

Gemini Interactions

Same idea for Gemini. You send OpenAI Chat Completions; we translate to the Interactions API (not generateContent) so implicit cache hits land in total_cached_tokens.

ts

const res = await client.chat.completions.create({
  model: "gemini-3.7-flash",
  messages: [{ role: "user", content: "Continue from last turn" }],
  // Optional: continue a stored Interactions thread for cache hits.
  previous_interaction_id: process.env.GEMINI_THREAD,
});

Upstream that becomes a POST to Google's Interactions API with store on by default and previous_interaction_id when you pass it. Thought tokens bill once as output. There is no public gateway /v1/interactions route.

Prompt caching

Stable prefix first, read usage.prompt_tokens_details.cached_tokens, billed at the cached-input rate. Optional prompt_cache_key. Details: prompt caching.

Long-context price breaks

GPT-5.6 flips the whole request to long rates at 272k prompt tokens (2× input, 1.5× output). grok-4.6 flips at 200k (2× input, 2× output, 2× cache). Those numbers are price breaks, not context windows.

Tools we reject

Caller-defined function tools pass through. Hosted tools that bill off the token meter are rejected: OpenAI web_search / file_search / code_interpreter, xAI web_search / x_search / code_execution, Gemini google_search / google_maps / url_context and the rest of that set. Use our POST /v1/search if you need live web results on the token meter.