You have a chat client or writing tool you like, but you want its model to run on your computer. If the client supports a custom OpenAI-compatible chat endpoint, OGAD (Off Grid AI Desktop) can supply the local model behind it.
This guide is for a client running on the same computer as OGAD. The core local API is free. Download the app and model first; the local chat request can then work without internet. The client may still use online features of its own.
Check what your client supports
The client must let you change the API base URL and model identifier. It must support chat requests through /v1/chat/completions. A client limited to one provider's account or to a different API cannot be made compatible just by entering a local address.
In OGAD, download and load a local text model from Models. Keep OGAD running. Check the gateway's address in the app; the usual port is 7878, but use the actual port if it differs.
Fill in the connection
Open your client's custom-provider or OpenAI-compatible connection settings. The labels vary, but the values have these roles:
| Client field | Value for the same computer |
|---|---|
| API base URL | http://127.0.0.1:7878/v1 |
| Full chat endpoint, if requested instead | http://127.0.0.1:7878/v1/chat/completions |
| Model | Exact local model ID returned by http://127.0.0.1:7878/v1/models
|
| API key | Leave blank if allowed; use a harmless placeholder if the client requires a nonempty field |
The gateway does not use that placeholder as authentication. Do not enter an unrelated provider's real key. Do not append /v1 twice when a client adds it itself.
Open the models URL in a browser and choose the id of a local chat model, not a remote entry or an image-generation model. Some clients discover this list automatically. Others need the ID pasted into their settings.
Get one useful result
Save the connection and use the client's test action if it has one. Then send a short request: “Rewrite this sentence in plain English: We will undertake a review of the proposed changes.”
Check that the selected provider is your local connection. If it succeeds, repeat with internet disconnected after the model is ready. That checks this route; it does not establish that every feature of the client is offline.
A 404 error often means the base path is wrong. Connection refused usually means OGAD is stopped, the port differs, or the gateway cannot be reached. A model error calls for a check of the returned model ID and its load state.
Keep the network boundary clear
The bundled gateway does not add an API-key login. Keep it on a trusted network; do not expose it to the public internet. If you later use a client on another device, 127.0.0.1 points to that other device, not your computer. That requires a separate reachable-address setup.
The endpoint is documented in the OGAD API reference. Open your compatible client's settings, point it at OGAD, and get one answer from the model you chose.
For further actions, you may consider blocking this person and/or reporting abuse