Use the model router

warpdrv isn’t just a chat app — it’s a model host. The model router exposes your running models through a standard OpenAI-compatible endpoint, so you can use them outside warpdrv: in a coding agent, your own editor, a custom script, or any OpenAI-compatible client.

Common harnesses people wire up to it include OpenCode, Cline, Roo Code, and Aider — plus any CLI or IDE agent that supports a custom OpenAI-compatible base URL.

Set up aliases

An alias is a short name you attach to a server. On the Servers page, click the + button to the right of a server’s name and type a short alias (e.g. qwen27b).

Why use an alias instead of the model filename?

Model filenames are long, unwieldy, and change every time you swap a quant or version. An alias is short, stable, and — most importantly — decouples the name your clients use from the actual model behind it.

This lets you put the same alias on two different servers and swap between them without touching your clients:

1. Set the alias qwen27b on a server running Qwen 3.6 27B.
2. Later, you want to try a bigger model. Add the same alias qwen27b to a second server running a different model.
3. Stop the first server and start the second.

Your OpenCode (or any other) config still says qwen27b — no changes needed — and it now talks to the new model automatically.

Sticky routes

When multiple servers share the same alias and are all running, the router assigns a sticky route: the first request picks one server (preferring a healthy one) and pins it, so all subsequent requests for that alias go to the same model. This keeps a conversation consistent instead of bouncing between servers mid-chat.

You can view and clear sticky routes, and start/stop the router, on the Proxy page.

Using it with a third-party app

This is the main use case. You’ll connect your local models to an external app by giving it two things:

• Your local IP address (or localhost if the app runs on the same machine).
• The port the router is running on (default 1234).

The base URL is:

http://<ip>:1234/v1

The alias is sent as the model field in each request — it is not part of the URL. So a client addressing the qwen27b alias sends {"model": "qwen27b", ...}.

OpenCode

OpenCode supports custom OpenAI-compatible providers. Add a local provider to your opencode.json pointing at the router:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "local",
      "options": {
        "baseURL": "http://localhost:1234/v1"
      },
      "models": {
        "qwen27b": {
          "name": "qwen3.6-27b",
          "limit": { "context": 260000, "input": 260000, "output": 260000 }
        }
      }
    }
  }
}

The key under models (qwen27b) is the alias you set on the server. For a remote machine, replace localhost with your local IP.

Other harnesses

Most OpenAI-compatible clients follow the same pattern — set a custom base URL to http://<ip>:1234/v1 and use your alias as the model name:

• Cline (VS Code) — set the OpenAI API Base to the router URL and the Model ID to your alias.
• Roo Code (VS Code) — same: custom base URL + model name.
• Aider — point --openai-base-url (or the OPENAI_BASE_URL env var) at the router and set --model to your alias.

If you use a different harness, look for its “OpenAI-compatible” / “custom base URL” setting and apply the same base URL + alias.

Security

The router is a network-accessible endpoint. If your machine’s IP is reachable beyond localhost (e.g. on a shared network), enable a Bearer token so only you can use it:

1. On the Proxy page, open the Access Tokens section.
2. Create a token and toggle Require Bearer token for /v1/*.
3. Pass that token as the API key in your client.

There’s also an option to enforce auth even for localhost requests if you want to test the protected path.

Next steps

• Improving your chat experience — get the most out of warpdrv’s own chat.
• Set up a workspace — organize projects and their default servers.