OpenAI’s Codex coding assistant can now skip its own models and talk to third-party AI providers—including open-source alternatives. A widely shared KuCoin tutorial claims you can start using DeepSeek V4 Pro or GLM-5.2 with one click. Reality is messier. Official Codex documentation doesn’t list those models as built-in integrations, and the recommended setup tool adds a translation layer you may not want to trust with your production code.

What the update actually delivers

Codex is a suite of coding tools OpenAI launched in mid-2026—desktop app, CLI, and SDK—all rewritten in Rust and open-sourced under Apache 2.0. The CLI and desktop app can act as a front-end for any model that speaks the right protocol, thanks to two mechanisms that shipped in July 2026.

First, the --oss command-line flag. When you start Codex with --oss, it bypasses OpenAI’s API and connects to a local model provider running on your machine. The official configuration guide names two first-party options: Ollama and LM Studio. You can set a default in ~/.codex/config.toml or pass a provider name at launch. That’s the extent of “official support” as of now.

Second, custom provider configuration. In that same config.toml file, you can define a provider block with a model name, base URL, authentication environment variable, API protocol (Responses or Chat Completions), and any extra headers. OpenAI’s docs show examples for a generic proxy, a local Ollama instance, and Mistral—not DeepSeek, not GLM-5.2. But any endpoint that mimics OpenAI’s API shape can be plugged in. That’s where the KuCoin article grabbed attention: it walked through using the third-party tool CC Switch to route Codex requests to DeepSeek or Zhipu AI’s GLM-5.2.

Those two models are not listed as named integrations in Codex’s own documentation. You can point Codex at them, just as you can point it at any REST API that accepts JSON. The distinction between “you can technically do this” and “OpenAI officially supports this” is what separates the KuCoin tutorial from what Codex admins will find when they open the official advanced config page.

Why Windows users should pay attention

If you’re a Windows developer who has been waiting for a local LLM coding partner, the --oss path with Ollama is the safest starting point. Ollama runs natively on Windows, installs like any other app, and downloads models with a single command. Launch Codex with --oss pointing at Ollama, select a model you’ve already pulled, and you’re working locally—no API keys changing hands. For home lab enthusiasts and anyone handling proprietary codebases, that’s a big deal.

IT administrators, however, need to read the fine print. The custom provider route—pointing Codex at a remote API like DeepSeek’s or Zhipu’s—requires an API key stored in an environment variable. That key, and every prompt and file snippet you send, leaves your machine and lands on someone else’s servers. Before swallowing a tutorial that glazes over this, verify:

  • Does the provider support the exact API protocol Codex expects? Codex uses OpenAI’s Responses API by default; many third parties only offer Chat Completions. The CC Switch tool mentioned in the KuCoin piece solves this by inserting a local proxy that translates between protocols. That proxy runs at 127.0.0.1:15721 and rewrites requests on the fly—meaning you now have another piece of software with full visibility into your code and prompts.
  • What features break? GitHub’s Codex issue tracker (#24659 and related threads) is already collecting reports about stream interruptions, missing model metadata, tool-call failures, and context-handling glitches when custom providers don’t implement every endpoint the agent workflow demands.
  • What does your compliance policy say? Sending source code to a third-party model endpoint may violate data-residency rules. Local models via Ollama stay local; remote providers do not.

How we got here: from closed fortress to open lobby

OpenAI spent years guarding model details so tightly that the community nicknamed it “ClosedAI.” GPT-4’s technical report in 2023 famously omitted architecture specifics “due to competitive and safety considerations.” In 2026, the same company released Codex’s source code under Apache 2.0 and published a guide telling you how to avoid paying for its own models.

This U-turn isn’t altruism. The model race is flattening. Benchmarks show GPT-5.5, Claude Opus, Gemini, and open-source contenders like DeepSeek V4 and GLM-5.2 all clustering near the top. When model capability is table stakes, the new battlefield is the toolchain. Codex is OpenAI’s play to become the command center from which developers launch any AI task. If you grow dependent on its CLI, its interface, its git-aware agent loops, the cost of switching to a competitor’s equivalent climbs—even if the actual LLM behind the curtain changes. Microsoft did the same with Windows; Google did it with Android. Codex, not any single model, is the long-term lock-in.

For Windows users, this shift also rides a larger trend: local AI tooling finally reaching maturity. Ollama for Windows arrived in 2024; LM Studio followed with GPU-accelerated inference on consumer hardware. Codex plugging into those runtimes means a developer with a gaming laptop can run a 7B–13B parameter model locally and still get IDE-grade coding assistance offline.

A practical setup for Windows, without the hype

Skip the CC Switch tutorial unless you’ve audited its code and accept the proxy risk. Here’s what actually works with what OpenAI documents today.

Option A: Use Ollama for local models

  1. Install Ollama from ollama.com.
  2. Pull a coding-capable model: ollama pull codellama:13b or ollama pull deepseek-coder:6.7b. (DeepSeek-Coder models run locally; this is not the same as calling DeepSeek’s cloud API.)
  3. In a terminal, launch Codex with codex --oss --oss-provider ollama --model codellama:13b.
    - To set a default, edit %USERPROFILE%\.codex\config.toml:
    toml [oss] default_provider = "ollama" default_model = "codellama:13b"
  4. Codex will load the model metadata and start the session. No API key needed; everything runs on your machine.

Option B: Point Codex at a custom provider (manual)

  1. Create or edit %USERPROFILE%\.codex\config.toml.
  2. Add a [providers.NAME] block. For a cloud provider that mimics OpenAI’s Chat Completions:
    toml [providers.deepseek] api = "chat" base_url = "https://api.deepseek.com/v1" env_key = "DEEPSEEK_API_KEY" [providers.deepseek.models] "deepseek-v4-pro" = {}
  3. Set the environment variable outside the config file—never paste the key into it:
    powershell
  4. Restart Codex and switch to the model with /model deepseek-v4-pro.
  5. Before you code anything real, run a smoke test on an empty throwaway repo. Verify that streaming, multi-turn conversations, and any agent commands (/fix, /explain) behave as expected.

Security checklist

  • 🔒 Keep approvals enabled so Codex won’t execute commands automatically.
  • 🧪 Use a disposable test repository for initial compatibility checks.
  • 🛡 Restrict Codex’s sandbox to project directories—don’t let it read your entire drive.
  • 🌐 Know where your data goes: --oss with Ollama is local; a custom provider configured with a cloud API key sends code off-device.
  • ⚠️ Third-party proxy tools are not endorsed by OpenAI. If you use CC Switch or similar, audit the code, block it from phoning home, and accept that it sits as a man-in-the-middle between Codex and your chosen model.

What comes next

OpenAI hasn’t announced official DeepSeek or GLM-5.2 integrations, but the custom-provider door is open. As more developers push on that door, expect either:
- Official partnerships that smooth out protocol mismatches, or
- Community-maintained provider definitions that harden the experience.

The bigger signal is that Codex will continue absorbing third-party models into its workflow. For Windows developers, the message is clear: you can pick your AI’s brain, but you’re still renting the skull from OpenAI.