Overview
Codex Model Switcher is a tool that connects compatible model providers to Codex (ChatGPT App), allowing users to select from providers such as DeepSeek, MiniMax, Zhipu GLM, Moonshot Kimi, Qwen, SiliconFlow, OpenRouter, Groq, a custom gateway/relay, or a local Ollama directly from the model dropdown in Codex, with a one-click way back to official OpenAI. It addresses the limitations of hand-editing the Codex configuration file, such as new Codex builds refusing certain wire APIs, requiring a hand-written model catalog JSON, storing API keys in plain text, and old conversations being tied to their original provider. The tool automates provider setup and switching, manages model catalogs, and ensures protocol compatibility.
The tool provides a one-command setup with 17 built-in presets or custom Base URL support, generates a full model catalog from the platform, and offers one-click switch/restore via CLI, interactive menu, or GUI. It preserves existing task bindings and raw history, never migrating them between providers. It includes a context guard that checks session size against the usable window and offers a one-click switch to a model that fits. Model capabilities are probed with real requests and written back into the catalog, and API keys are stored securely in the system keychain.
Security features include atomic config writes with concurrent-write protection, masked keys in UI/logs, and local-only GUI and bridge listening on 127.0.0.1. The tool supports macOS with a native app and Windows/Linux with a browser UI, and has a bilingual UI (Chinese and English) with zero dependencies. It also auto-wires the provider's API surface, including MCP servers and environment files for official CLIs, SDKs, and curl.
Pricing
Free
Detailed plans have not been confirmed in our catalog. Check the official website for current limits and billing terms.
Visit WebsitePrices and limits may change. Confirm the currency, billing period, seat minimum and usage caps on the official website.
Use Cases
- Developers use Codex Model Switcher to connect third-party model providers like DeepSeek or MiniMax to Codex without manually editing configuration files.
- Teams use the tool to switch between model providers for new tasks while keeping existing task bindings and history intact.
- Administrators use the tool to securely store API keys in the system keychain instead of plain text in config files.
- Developers use the tool to auto-wire a provider's API surface, including MCP servers and environment files, for use with official CLIs, SDKs, and curl.
- Users use the tool to restore official OpenAI configuration with a single command, keeping third-party config intact.
- Teams use the tool to diagnose and fix common issues like unknown model errors or protocol mismatches via built-in diagnostics.
Features
Connect multiple model providers to Codex
Switch Codex models with one click
Use a native macOS app
Run from Windows and Linux terminals
Check model account balances
View usage as percentages
Bridge the Responses protocol
Run without third party dependencies
Reviews
No reviews yet. Be the first to share your experience!
FAQ
Does continuing an old conversation after switching work?
Yes for third-party to third-party: recent tasks are re-bound automatically. Threads bound to official OpenAI are intentionally left alone (they'd lose their account binding) — fork or start a new task for those.
What about image generation / web search / computer use?
Not available on third-party providers. Those are OpenAI server-side tools or separate platform models (image-01, video-01…); Codex talks to one chat model. The tool won't pretend otherwise — that's exactly what the capabilities view is for.
Why two "count the squares" images for the vision probe?
Single-image color questions get guessed correctly far too often — the probe was fooled once. Two different counts must both be right, and the output budget is generous so reasoning models don't return an empty (truncated) answer and get misread as "no vision".
Does it touch my plugins / MCP / skills?
No. Model-related fields only, with verify-and-rollback.