Choose a client

Every client needs the same two things: the gateway's address and one of the keys in your api_keys. Nothing is installed into the client.

Client	API	Base URL	Guide
Codex CLI	OpenAI Responses	http://127.0.0.1:8787/v1	Codex CLI
Claude Code	Anthropic Messages	http://127.0.0.1:8787	Claude Code
OpenCode	Anthropic Messages	http://127.0.0.1:8787/v1	OpenCode
Pi	OpenAI Responses	http://127.0.0.1:8787/v1	Pi
Crush	Anthropic Messages	http://127.0.0.1:8787	Crush
Zuno	OpenAI Responses	http://127.0.0.1:8787/v1	Zuno
OpenAI SDKs	OpenAI Responses	http://127.0.0.1:8787/v1	below
Anthropic SDKs	Anthropic Messages	http://127.0.0.1:8787	below

The base URL depends on what the client appends to it. Claude Code, Crush and the Anthropic SDKs add /v1/messages; OpenCode's @ai-sdk/anthropic adds /messages only, so its base URL keeps /v1.

A newer client release can change the requests it sends; if a field it adds is not supported yet, the gateway names it in a typed error instead of ignoring it.

The key

The gateway accepts the key as Authorization: Bearer <key> or as x-api-key: <key> on every route, so both SDK families work unchanged. Each key is a tenant of its own: stored responses, session bindings and reasoning history made with one key are invisible to requests with another. Giving each client its own key keeps them apart; a conversation has to continue with the key that started it.

Model names

GET /v1/models lists the models your accounts can use. Besides the plain names, such as gpt-5.6-sol or claude-opus-5-5, it lists auto and one name per effort level for models that have levels, such as claude-opus-5-5-high. Codex CLI reads the effort levels from the same list for its /model picker; the Codex guide shows how to load it.

Switching model or effort in the middle of a conversation works: the visible history and tool calls carry over, and reasoning that the new model cannot read is left out and reported in a response header. The configuration reference explains the rules.

SDKs

The OpenAI SDK for JavaScript:

import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: "http://127.0.0.1:8787/v1",
  apiKey: process.env.KIRO_GATEWAY_API_KEY,
});

const response = await openai.responses.create({
  model: "gpt-5.6-sol",
  store: false,
  input: "Reply with exactly: KIRO_OK",
});
console.log(response.output_text);

The Anthropic SDK for JavaScript. Its base URL has no /v1, because the SDK adds it:

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

const anthropic = new Anthropic({
  baseURL: "http://127.0.0.1:8787",
  apiKey: process.env.KIRO_GATEWAY_API_KEY,
});

const message = await anthropic.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Reply with exactly: KIRO_OK" }],
});
console.log(message.content[0].type === "text" ? message.content[0].text : "");

kiro-provider passes max_tokens on to Kiro for claude-opus-5-5, claude-opus-5, claude-sonnet-5 and claude-fable-5-1, from 1,024 to 128,000; a value outside that range is refused rather than changed. Kiro has no output limit kiro-provider can set for other models, and every Messages request carries max_tokens, so Messages serves those four models, plus the GPT-5.6 models when the request has the header x-kiro-output-token-limit-mode: advisory (OpenCode sets it). On Responses, max_output_tokens follows the same rule but is optional; a request without it can use every model.

Separate commands for Kiro

The client launchers set up a kirocodex and a kiroclaude command next to the codex and claude you already have, each with its own home directory, so your usual sessions and settings stay as they are.

Other clients

Any client that speaks OpenAI Responses or Anthropic Messages can try the same base URLs. The protocol compatibility reference lists which request fields are supported. Clients that only speak Chat Completions need the route turned on with enable_legacy_chat_completions; it does not carry Responses session metadata.