codex-model-router

# codex-model-router

Use models from **different providers in the same OpenAI Codex setup** — switch between GPT and third-party models (e.g. Meta Muse Spark) just by picking a model name, with no `config.toml` edits and no app patching.

A ~90-line, zero-dependency Node.js forwarder on `127.0.0.1`. Codex talks to it as its normal OpenAI endpoint; the router reads the `"model"` field of each request and forwards it, unchanged, to the matching provider with that provider's API key. Responses stream straight through.

```text
Codex ──► http://127.0.0.1:4141/v1 (router) ──┬─ "muse-*"  ─► api.meta.ai/v1      (META_API_KEY)
                                              └─ anything ─► api.openai.com/v1   (your Codex OpenAI key, passed through)
```

Context: [openai/codex#46484](https://github.com/openai/codex/issues/46484) (native multi-provider model switching). Today `model_provider` is global, so a custom provider and GPT models can't coexist in one model picker. This router is a no-patch workaround.

## One-click install (Windows)

Prerequisites: [Node.js](https://nodejs.org) in `PATH`, and your provider keys as user environment variables, e.g.:

```powershell
[Environment]::SetEnvironmentVariable('META_API_KEY', '<your-meta-key>', 'User')
```

Then double-click **`install.cmd`** (or run `powershell -ExecutionPolicy Bypass -File install.ps1`). It:

1. copies the router to `%LOCALAPPDATA%\codex-model-router`
2. adds a hidden launcher `codex-model-router.vbs` to your Startup folder (runs at login)
3. backs up `~/.codex/config.toml`, then sets at the top:
   ```toml
   model_provider = "openai"
   openai_base_url = "http://127.0.0.1:4141/v1"   # codex-model-router
   ```
4. starts the router and checks `http://127.0.0.1:4141/health`

Restart the Codex app afterwards. Undo everything with `uninstall.ps1`.
Options: `-Port 4141 -CodexHome <dir> -InstallDir <dir> -StartupDir <dir> -NoStart`.

macOS/Linux: no installer yet — run `node codex-router.js` (e.g. via launchd/systemd) and add the two config lines above.

## Adding a provider or model

**1. Route** — edit `routes.json` (in the install dir). First matching `match` prefix wins; `""` is the catch-all.

```json
[
  { "match": "muse-", "host": "api.meta.ai", "base": "/v1", "keyEnv": "META_API_KEY" },
  { "match": "",      "host": "api.openai.com", "base": "/v1", "keyEnv": "OPENAI_API_KEY", "passthroughAuth": true