OpenAI's coding agent Codex used to accept only its own GPT models — now it has opened up the "model access layer." Through a config called model_providers, you can point Codex at local model services like Ollama and LM Studio, or even third-party APIs like DeepSeek. That means you can switch models on demand, save tokens, and even run the coding agent fully offline.

What Codex open-source mode is

Essentially, what OpenAI opened up here isn't the model itself but a "pluggable model interface layer" for Codex. You register multiple model providers in the config file, and Codex picks the corresponding provider at startup based on the config, routing requests to different model services.

Core features

The --oss switch

The most direct usage: add a --oss flag on the command line and Codex connects straight to a local open-source model service. Two are supported out of the box:

• Ollama: the most popular tool for running large models locally
• LM Studio: a desktop alternative with a graphical interface

The model_providers config

The more flexible route is registering model providers in the config file; each provider carries four pieces of information:

• base_url: the access address
• wire_api: the wire protocol (currently only responses is accepted)
• env_key: the authentication method
• model: the model mapping

Mistral, enterprise-internal proxies, and third-party relay stations can all connect this way. You can also save settings as config profiles and switch between them by name on the command line while debugging.

In practice: --oss with a local model

With a local model service and the network permission settings in place, you can have Codex do code generation and reasoning entirely on your machine, achieving offline runs and local processing.

The real-world experience

Advantages

• No single-vendor lock-in: switch models on demand — privacy and cost are your call
• Saves money: let GPT handle the planning (breaking down tasks, designing architecture) while open-source models handle execution (writing code, batch-editing files) — the same task can cost a fraction as much
• Local and offline: not a single line of code leaves your computer; a fit for individual developers who don't want their private projects uploaded to the cloud

Limitations

An installed socket doesn't mean every appliance you plug in will run. Incoming models generally need to be compatible with the Chat Completions interface format; as for whether more complex capabilities like function calling work end to end, the official team makes no guarantees — you'll have to test them one by one.

How to connect DeepSeek

DeepSeek is one of the open-source models most familiar to many Chinese developers. But pasting DeepSeek's address straight into Codex doesn't work smoothly — new Codex builds are based primarily on OpenAI's Responses API, while DeepSeek and most open-source model endpoints still speak Chat Completions.

The community's fix is to insert a local "routing layer" or "protocol converter" in the middle. The basic flow:

1. Codex sends requests per the Responses API
2. The routing layer converts them into Chat Completions format
3. They're forwarded to DeepSeek and other open-source models
4. The returned results are converted back into the Responses format Codex can recognize

Similar capability isn't limited to CC Switch. LiteLLM, claude-code-router, and the various proxy services developers build themselves are all, in essence, solving the same problem.

Use cases

• Individual developers: run Codex on local open-source models — cheaper and privacy-preserving
• Hybrid routing: GPT plans + open-source models execute, cutting costs roughly in half
• Enterprise internal deployment: connect enterprise-internal proxies via model_providers and manage model access in one place

💡 Tip: what OpenAI opened up here is the "model access layer," not the models themselves. Whichever model you connect, it has to align with OpenAI's defined request and response structures. This interface protocol is becoming the new competitive battleground for AI coding tools.