Run OpenAI Codex on Mingles Router
Codex is OpenAI's own coding agent — it reads your repo, runs commands and edits files from the terminal. It also accepts a custom provider, so the model behind it can be ours. One block in config.toml and it runs on DeepSeek-V4-Flash or Kimi-K2.6 instead of GPT.
Pick your model and paste your key — every command and value on this page updates live.
Endpoint: https://router.mingles.ai/v1 · key stays in your browser ·
no key entered? commands show <your-key>.
Install + configure
1. Install Codex
npm i -g @openai/codexVerified against codex-cli 0.147.0.
2. Add the provider block
Create or edit ~/.codex/config.toml. The two fields people get wrong are wire_api and requires_openai_auth — see below.
model = "deepseek-ai/DeepSeek-V4-Flash-0731"
model_provider = "mingles"
[model_providers.mingles]
name = "Mingles Router"
base_url = "https://router.mingles.ai/v1"
env_key = "MINGLES_API_KEY"
wire_api = "responses"
requires_openai_auth = false3. Export your key and run it
export MINGLES_API_KEY=<your-key>
codexCheck the wiring first with `codex doctor` — it prints the resolved provider, the model, and whether it found your key.
Exact values
| Config file | ~/.codex/config.toml |
| base_url | https://router.mingles.ai/v1 |
| wire_api | responses |
| env_key | MINGLES_API_KEY |
| model | deepseek-ai/DeepSeek-V4-Flash-0731 |
Switch model
Already using Codex on an OpenAI account? Nothing is overwritten — model_provider selects which block is live, so you can keep both and switch by editing that one line.
Before you file a bug: read the limits
Reasoning models spend the output budget on internal thinking, output is capped at 8192 tokens, there is no KV cache, and there are no built-in web tools. Most “it broke on Mingles Router” reports are one of these. See Model limits & behavior →
Frequently asked
Why wire_api = "responses" and not "chat"? +
Because "chat" no longer exists. Codex removed it — set it and you get `wire_api = "chat" is no longer supported`. Codex now speaks only the Responses API, so we implemented /v1/responses on top of our chat-completions pipeline. That is what this config talks to; the rest of our API is unchanged.
Does the agent part work, or just chat? +
The agent part. We tested codex-cli 0.147.0 against production with `--sandbox workspace-write`: it called a shell tool, created a file, ran it, read the output and reported back. Tool calls, streaming and multi-turn tool results all round-trip.
“Model metadata for … not found. Defaulting to fallback metadata”? +
Harmless. Codex ships a table of known OpenAI models and warns for anything outside it. It affects Codex’s own context-window bookkeeping, not the request. Everything works with the warning present.
I see raw JSON with "reasoning" and "tool_calls" in the transcript +
A known rough edge on our side, not a Codex bug. Where a model’s tool calls are emulated, the streaming path currently emits the emulation JSON as visible text as well as a proper tool call. The tool still runs correctly. It is on our list.
Do I need an OpenAI account or API key? +
No. `requires_openai_auth = false` tells Codex to use your provider’s key instead of ChatGPT sign-in. Only MINGLES_API_KEY is needed.
Codex or DeepSeek Harness? +
Codex is the mature one and runs in your terminal; DeepSeek Harness is MIT-licensed, runs as a local web UI and is in developer preview. Both take our endpoint, so it is a question of which interface you want.