Use apimodels in OpenAI Codex

Fastest: the apimodels plugin for video and images

Codex does the planning — script, shot list, prompts, editing with ffmpeg; the plugin renders the video, images and voice-over. It ships with a skill, so Codex decides on its own when to use it, which model to pick, how long video takes and when to ask you before a batch — just ask in plain words. The plugin does not change the model Codex runs on; it works with a ChatGPT sign-in too.

npx -y apimodels-mcp login

codex plugin marketplace add stimQQ/apimodels-mcp
codex plugin add apimodels@apimodels

3. Just ask, e.g. "make a 5-second 16:9 video of a kitten slowly raising its head in morning light". Generating video, images or speech spends credits, so Codex asks you to confirm each call; checking progress or the balance never asks.

Node.js is required. If it is missing, install the plugin anyway and ask for something: Codex installs the official Node.js for you once you agree (on Windows click Yes in the system prompt; on macOS click Continue, then Install), then restart Codex. Codex can also run the sign-in step for you; you only click Authorize in the browser.

Get running in 3 steps

1. 1

   Grab an API key

   Open Console in your console and create an sk_… key.

2. 2

   Drop the config

   Paste the snippet below into ~/.codex/config.toml (create it if needed; on Windows it lives in C:\Users\<you>\.codex\).

3. 3

   Set the key + start

   Put your key in the env var, then (re)open Codex.

Full config

# ~/.codex/config.toml — drop-in, paste as-is
model_provider           = "apimodels"
model                    = "gpt-6-sol"        # or "gpt-6-luna" (lowest-priced) / "gpt-5.5"
review_model             = "gpt-6-sol"        # model used by /review
model_reasoning_effort   = "medium"           # low | medium | high
disable_response_storage = true

[model_providers.apimodels]
name     = "apimodels"
base_url = "https://api.apimodels.app/v1"
wire_api = "responses"
env_key  = "APIMODELS_API_KEY"

# put your apimodels key in the env var the config points at,
# then launch codex (add this line to ~/.zshrc to make it stick):
export APIMODELS_API_KEY="sk_…your_key…"
codex

# Windows (PowerShell or cmd) — then reopen Codex:
setx APIMODELS_API_KEY "sk_…your_key…"

Settings reference

Which models Codex can use

The model setting is the model Codex itself uses to talk, read and edit code, and every turn goes through /v1/responses. So it only takes chat models on that endpoint; anything else returns HTTP 400 with the reason in the error message.

Without the plugin: configure the MCP server by hand

Video and image models can't go in model, but Codex can call them as tools. Without the plugin, add our MCP server, apimodels-mcp, to the same config.toml and Codex gains generate_video, generate_image, text_to_speech and more, covering every video and image model on the site. Node.js must be installed (it runs through npx). After the login command above, the env line can go.

# add to the same ~/.codex/config.toml
[mcp_servers.apimodels]
command = "npx"
args    = ["-y", "apimodels-mcp"]
env     = { APIMODELS_API_KEY = "sk_…your_key…" }
# tool_timeout_sec = 660   # optional: lets one call wait for the whole video

Restart Codex and ask. Without the plugin's skill Codex may not go looking for these tools, so name the tool in your request:

Use apimodels generate_video with grok-imagine-video-1.5 to make a 10-second 16:9 video:
a golden tabby kitten slowly raises its head in warm morning light.

Video takes a while: a median of about 2.5 minutes, 9 in 10 within 8 minutes. Codex aborts a tool call after 60 seconds by default, so each call waits at most 50 seconds and then hands back a task id; Codex calls get_task on its own until it has the video URL. Nothing is resubmitted or billed twice.

On Windows, if the MCP server does not start, launch it through cmd:

# Windows: if the MCP server does not start, launch npx through cmd
[mcp_servers.apimodels]
command = "cmd"
args    = ["/c", "npx", "-y", "apimodels-mcp"]
env     = { APIMODELS_API_KEY = "sk_…your_key…" }

All tools, parameters and env vars are in /docs/mcp.

Reasoning effort

Reasoning depth is the reasoning.effort field in the request body (Codex sets it via model_reasoning_effort), not a model-name suffix. All levels share the same per-token price; higher effort just emits more reasoning_tokens (billed as part of output_tokens). When calling the API directly, pass "reasoning": { "effort": "high" }.

30-second sanity check

Before setting up Codex, confirm the endpoint and your key work with a single curl:

curl -s https://api.apimodels.app/v1/responses \
  -H "Authorization: Bearer $APIMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "input": "Reply with exactly: ok",
    "reasoning": { "effort": "low" },
    "max_output_tokens": 16
  }'

Expected response (truncated):

{
  "id": "resp_...",
  "object": "response",
  "status": "completed",
  "output": [{
    "type": "message",
    "role": "assistant",
    "content": [{ "type": "output_text", "text": "ok" }]
  }],
  "usage": { "input_tokens": 22, "output_tokens": 5, "total_tokens": 27 }
}

HTTP 200 with output[].content[0].text === "ok" means you're good to go.

responses or chat?

Codex uses wire_api = "responses" — its native mode, which handles reasoning and multi-turn tool calls most cleanly — so pick responses for Codex. That also sets which models Codex can use: Claude and Gemini are only on /v1/chat/completions, and work as usual in clients that speak chat format (Cherry Studio, Cursor and the like).

Troubleshooting