How to Connect Third‑Party LLMs to Codex App, CLI, and SDK

Provider configuration basics

Codex App, CLI, and SDK can connect to any LLM that implements a compatible OpenAI‑style API. The connection details are defined in a provider block placed in the user‑level configuration file ~/.codex/config.toml. Project‑local .codex/config.toml files are ignored.

A provider block consists of four keys: name – human‑readable identifier base_url – HTTP endpoint of the model API env_key – name of the environment variable that holds the API key wire_api – protocol to use ("responses" or the legacy "chat")

Reserved provider IDs openai, ollama, lmstudio, and amazon‑bedrock cannot be reused for custom providers.

Example: Adding Qwen (DashScope) as a provider

[model_providers.qwen]
name = "Qwen 通义千问 (DashScope)"
base_url = "https://dashscope.aliyuncs.com/api/v2/apps/protocols/compatible-mode/v1"
env_key = "DASHSCOPE_API_KEY"
wire_api = "responses"

After saving the file, export the API key:

export DASHSCOPE_API_KEY="<YOUR_KEY>"

To select the model without editing the file, override the configuration on the command line:

codex -c model='"qwen3.7-max"' -c model_provider='"qwen"'

Wire API selection

The wire_api field tells Codex which protocol to use. Earlier versions accepted chat, which maps to the OpenAI /v1/chat/completions endpoint. Starting with codex‑cli 0.140.0 , chat is treated as a hard error. The recommended value is responses, which maps to the newer /v1/responses endpoint.

Models that already support the Responses API

Qwen (DashScope) – use the /v2/apps/protocols/compatible-mode/v1 path.

MiniMax – supports POST /v1/responses.

Step models from 阶跃星辰 – also expose the Responses endpoint.

model = "MiniMax-M3"
model_provider = "minimax"
model_context_window = 512000

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimaxi.com/v1"
experimental_bearer_token = "<MINIMAX_API_KEY>"
wire_api = "responses"

Models that only expose Chat Completions

DeepSeek, GLM, Kimi, and Xiaomi MiMo currently provide only /chat/completions or Anthropic‑compatible endpoints. To use them with Codex you must either:

Set wire_api = "responses" and run a proxy that translates Responses requests to Chat Completions, or

Employ an existing bridge such as cc-switch, mimo2codex, or va-ai-api-bridge.

Local model providers

Codex includes built‑in support for local providers via Ollama and LM‑Studio. Launch a local model without writing a provider block by using the --oss flag:

codex --oss-m <MODEL_NAME>

Or set the default local provider in config.toml:

oss_provider = "ollama"  # or "lmstudio"

Ollama (v0.13.3+) and LM‑Studio both implement the OpenAI‑compatible POST /v1/responses endpoint, allowing direct use with Codex as long as the host machine has sufficient resources.

Key points

Add a third‑party LLM by creating a provider entry in ~/.codex/config.toml with the appropriate wire_api value.

Prefer responses over the deprecated chat protocol (codex‑cli 0.140.0 enforces this).

Verify that the target model supports the Responses API; otherwise use a translation proxy or bridge.

Local models can be accessed directly via the --oss flag or by setting oss_provider to ollama or lmstudio.