Responses API and Codex

Codex and the OpenAI SDKs’ responses.create run on Runcrate models: point them at https://api.runcrate.ai/v1 with your Runcrate API key and a Runcrate model id. Requests are billed, rate limited and routed exactly as chat completions are, at the same per-token prices. Nothing is stored: every request carries its whole conversation, as Codex sends it.

Codex

Add a Runcrate provider to ~/.codex/config.toml:

model = "zai-org/GLM-5.3"
model_provider = "runcrate"

[model_providers.runcrate]
name = "Runcrate"
base_url = "https://api.runcrate.ai/v1"
env_key = "RUNCRATE_API_KEY"
wire_api = "responses"

Then run Codex with your key in RUNCRATE_API_KEY:

export RUNCRATE_API_KEY=rc_live_YOUR_API_KEY
codex

Any text model with tool calling in the model catalog works. Codex warns that it has no metadata for a model id it does not know and uses its defaults for the context window. Codex’s web search tool runs at OpenAI, so it is left out here: the model answers without it.

Endpoint

POST https://api.runcrate.ai/v1/responses

Basic usage

curl https://api.runcrate.ai/v1/responses \
  -H "Authorization: Bearer $RUNCRATE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3",
    "instructions": "Answer in one paragraph.",
    "input": "Explain quantum computing in simple terms",
    "max_output_tokens": 512,
    "store": false
  }'

from openai import OpenAI

client = OpenAI(base_url="https://api.runcrate.ai/v1", api_key="rc_live_YOUR_API_KEY")

response = client.responses.create(
    model="zai-org/GLM-5.3",
    input="Explain quantum computing in simple terms",
    store=False,
)
print(response.output_text)

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://api.runcrate.ai/v1", apiKey: "rc_live_YOUR_API_KEY" });

const response = await client.responses.create({
  model: "zai-org/GLM-5.3",
  input: "Explain quantum computing in simple terms",
  store: false,
});
console.log(response.output_text);

Streaming

With stream: true the answer arrives as response.* events: response.created and response.in_progress, then each output item as response.output_item.added, its parts and deltas (response.output_text.delta, response.reasoning_summary_text.delta, response.function_call_arguments.delta) and response.output_item.done, then response.completed with the whole response and its usage. A response cut short by max_output_tokens ends with response.incomplete; one that fails, with response.failed.

const stream = client.responses.stream({ model: "zai-org/GLM-5.3", input: "Write a poem", store: false });
stream.on("response.output_text.delta", (event) => process.stdout.write(event.delta));
const final = await stream.finalResponse();

Tools

Function tools take name, description and parameters. A call comes back as a function_call item; send it back in input with a function_call_output item carrying its call_id.

const tools = [{
  type: "function",
  name: "get_weather",
  description: "Weather for a city",
  parameters: { type: "object", properties: { city: { type: "string" } }, required: ["city"] },
}];
const ask = { model: "zai-org/GLM-5.3", tools, store: false, input: [{ role: "user", content: "Weather in Paris?" }] };

const first = await client.responses.create(ask);
const call = first.output.find((item) => item.type === "function_call");

const answer = await client.responses.create({
  ...ask,
  input: [...ask.input, ...first.output, { type: "function_call_output", call_id: call.call_id, output: "18C and sunny" }],
});

What is supported

Responses API	On Runcrate
`instructions`, and `developer` or `system` messages	Supported, as the model’s system prompt
`input` as a string, or `message`, `function_call`, `function_call_output`, `custom_tool_call` and `custom_tool_call_output` items	Supported
`input_text` and `input_image` (a URL or a data URL) parts	Supported. Images need a vision model.
`reasoning` items sent back	Accepted; not passed to the model again
`function`, `custom` and `namespace` tools	Supported
`tool_choice` (`auto`, `none`, `required`, a named tool), `parallel_tool_calls`	Supported, where the model supports forcing a tool
`max_output_tokens`, `temperature`, `top_p`, `prompt_cache_key`, `stream`	Supported
`reasoning.effort`	`minimal` and `low` are low, `medium` medium, `high` and `xhigh` high
`text.format` `json_object` or `json_schema`	Supported, where the model supports structured output
`store`, `include`, `metadata`	Accepted; nothing is stored, and the response says `store: false`
Hosted tools (`web_search`, `file_search`, `code_interpreter`, …)	Left out; the model answers without them
`previous_response_id`, `conversation`, `background`, `input_file`, image `file_id`	Not supported: 400

A model’s reasoning comes back as a reasoning item whose summary is the reasoning text.

Errors

Errors use OpenAI’s shape, {"error": {"message", "type", "code", "param"}}, with the same codes as chat completions, including the free allowance’s (free_allowance_used, free_tier_paused, free_capacity_busy, free_prompt_too_large, github_link_required, with its free_tier details beside them). A Retry-After header says when to retry a 429 or 503.

Welcome

SDKs

CLI

Inference

Compute

Jobs

Schedules

Secrets

Storage

Dedicated Clusters

Billing

Account

Webhooks

FAQ

Responses API and Codex

Codex

Endpoint

Basic usage

Streaming

Tools

What is supported

Errors

​Codex

​Endpoint

​Basic usage

​Streaming

​Tools

​What is supported

​Errors

Codex

Endpoint

Basic usage

Streaming

Tools

What is supported

Errors