Configure a model and a harness
Goal: connect one model service, make a model available to users, and check that a harness (runtime) is ready.
Before you start: you can sign in to the admin console and the client (Connect the desktop client). You have the address and API key of a model service you are allowed to use.
A model is the AI that reasons and writes. A harness (called Runtime in the client) is the program that uses the model to read files, run commands and call tools. You need both.
1. Connect a model service
In the admin console, open Model catalog and choose Connect a service.
| Field | What to enter |
|---|---|
| Identifier | A short English ID, for example my-gateway. It can't be changed later |
| Name | A display name, for example Engineering gateway |
| API protocol | OpenAI compatible. This works for most gateways, vLLM, Ollama, proxies and compatible providers |
| API address | The base URL including /v1, for example https://gateway.example.com/v1 |
| API Key | The organization key for this service |
| Description | Optional |
Choose Save.
The key is shared by the organization. It is stored on the Hub and is never returned in plain text. Users can also enter a personal key in the client; when they do, their personal quota is used.
The catalog also lists common providers as built-in services. To use one, open it, enter the organization key and enable the models you need.
2. Add a model
On the service card, choose Add model.
| Field | What to enter |
|---|---|
| Model purpose | Chat model for conversations and agents |
| Model ID | The exact name the gateway expects, for example deepseek-chat |
| Display name | Optional. Shown to users instead of the ID |
Save, and make sure the model is enabled. Repeat for each model you want to offer.
3. Check the model in the client
In the client, open Settings → My models. The service appears under In use with Available from company. Choose Test capabilities next to a model to check image input, tool calling and streaming.
4. Check the harness
Open Settings → Agent management.
- Pi is built in and shows Ready to dispatch.
- Other harnesses show Install or Not installed. Install them when you need them.
- Choose Test on a harness to run a short live check with your model.
| Harness | Model it uses |
|---|---|
| Pi, OpenCode, DeepSeek Harness, Qwen Code, Kimi Code, Codex Harness, Harness² | Models from your SoduWork catalog |
| Codex | Its own OpenAI account. Choose Connect Codex to sign in |
| Claude Code | Its own Anthropic account |
To add another program that supports ACP, see Add an ACP harness.
Check that it worked
- In a new chat, the model menu lists your model.
- The Runtime menu shows Pi as selected.