Codex DeepSeek Switcher

Switch OpenAI / DeepSeek in VS Code Codex without repeatedly editing provider settings. Keep the normal Codex profile separate from a DeepSeek profile on Windows or a Linux Remote SSH account.

Website · Released VSIX · Marketplace · 中文 · Troubleshooting

What changes — and what does not

Windows account                     Linux SSH account (separate home)
  OpenAI   → .codex                    OpenAI   → .codex
  DeepSeek → .codex-vscode-deepseek     DeepSeek → .codex-vscode-deepseek
  one saved provider choice           a separate saved provider choice

This is a configuration illustration, not a live API demonstration. Windows sharing the same OS account/home share the selection. SSH aliases or NFS mounts sharing a home also share state. Reloading applies the selection to new processes; running tasks are not hot-swapped.

Common setup questions

How do I use DeepSeek in the official Codex VS Code extension?

Install Codex DeepSeek Switcher (extension ID dzr.codex-local-model-switcher) alongside the OpenAI Codex extension (openai.chatgpt) in the same extension host. Initialize your normal Codex configuration, then run Codex Source: Use DeepSeek Direct. Enter your own DeepSeek API key in the masked prompt when requested, save your work and reload the window. This is an independent companion extension, not an official OpenAI or DeepSeek product. Step-by-step guide.

Can I keep my GPT / ChatGPT-authenticated Codex setup while using DeepSeek?

The switcher selects the normal .codex home for OpenAI and .codex-vscode-deepseek for DeepSeek. It does not copy the normal auth.json into the DeepSeek home. Windows using the same OS account and home share one provider choice; this is not per-window isolation. Running processes keep their original environment until restarted. Custom CODEX_HOME layouts are not preserved by these launchers. Step-by-step guide.

Does using DeepSeek with Codex require MoonBridge or a local proxy?

This extension's direct mode uses https://api.deepseek.com/ with wire_api = responses; no MoonBridge or localhost relay is required. It selects provider profiles rather than translating API protocols. It does not supply a DeepSeek account, API credits, model permissions or network connectivity. A ChatGPT login is not used to authenticate the DeepSeek profile. Step-by-step guide.

Where do I install the switcher for VS Code Remote SSH?

Connect to the Linux SSH host and install both openai.chatgpt and dzr.codex-local-model-switcher under SSH: your-host, not only under Local. The remote side needs its own Codex configuration, Bash, Python 3 and GNU utilities. A separate remote home has a separate provider choice; two SSH aliases or NFS mounts sharing one home do not create isolation. Step-by-step guide.

Install

Install the currently published Marketplace version:

code --install-extension dzr.codex-local-model-switcher

For a newer VSIX obtained from this repository's releases or a verified maintainer build, use Extensions: Install from VSIX... and select the actual downloaded file. The extension ID remains dzr.codex-local-model-switcher. GitHub source changes, build artifacts, releases and Marketplace updates are separate; consult dated publication evidence rather than assuming the channels match.

For Remote SSH, connect first and install under SSH: your-host, not only under Local. Install the OpenAI Codex extension (openai.chatgpt) on the same extension host. The Remote SSH guide explains the distinction.

Older VSIX builds identify as hanzaidao.codex-local-model-switcher: disable that old extension before enabling dzr.codex-local-model-switcher, so two switchers do not run together. Neither installing a new ID nor switching back automatically removes old keys or backups. See distribution notes.

Switch in three steps

1. In a disposable folder, confirm your existing OpenAI Codex works. Save work and privately back up your configuration.
2. Run Codex Source: Use DeepSeek Direct. Enter your own DeepSeek key through the masked prompt when required, then choose Reload current window when ready.
3. Run Codex Source: Show Current Codex Source Status to inspect configuration. A minimal request such as “Reply OK without using tools” separately checks connectivity and may incur provider charges.

To return, run Codex Source: Use GPT (OpenAI) and reload again. The status-bar entry opens the switcher. Codex Source: Reapply Current Codex Source Setup repairs the current setup; it is not an uninstall/reset command.

Requirements and limits

An empty config.toml is sufficient only when Codex defaults suit your setup; initialize Codex first. The launchers look for the binary bundled with the Codex extension, not just any codex on PATH. Custom CODEX_HOME layouts are not preserved by these launchers. See verification scope: automated fixture tests do not establish successful live model calls or every client-version combination.

Provider and security boundaries

DeepSeek uses https://api.deepseek.com/ and wire_api = "responses", with the bundled deepseek-v4-pro / deepseek-v4-flash catalog. No MoonBridge or local relay is required. Provider availability, model entitlement, network access and billing must be checked separately.

The normal .codex/auth.json is not copied into the DeepSeek home. However, the generated configuration can preserve unrelated provider/MCP values from the source configuration. The DeepSeek key is plaintext in the isolated config.toml, not encrypted by the masked input. Remote files use 0600; Windows protection relies on filesystem ACLs. Switching back does not erase credentials or retained backups.

Setup changes chatgpt.cliExecutable to codex-vscode-profile; Linux setup also adjusts the VS Code Server setting extensions.supportNodeGlobalNavigator. Activation may repair an already installed launcher. Read Privacy and Security before use. Never post authentication files, complete configs, private code, host addresses or raw logs in issues.

Guides and development

Use DeepSeek · Return to OpenAI · Remote SSH · Release checks

npm run verify
npm run test:remote    # Linux; Python 3.11+ is needed for the test suite
npm run build:site
npm run check:site
npm run package:list
npm run package:vsix

Packaging uses the exact tool version in package.json; no runtime dependency is added. CI checks Windows/Linux JavaScript tests, actual Windows launcher compilation/execution against a probe, Linux helper integration, static links and VSIX contents. The website is built locally into _site/; a successful build is not proof of a live website.

Source and MIT license are available in this repository. A star is welcome when the tool is useful; there are no star prompts in the extension or star-gated features. Statistics collection is an optional maintainer command, not user telemetry.

Independent community project. Not affiliated with or endorsed by OpenAI, DeepSeek, Microsoft or Visual Studio Code.