Models & providers
Connect providers, choose the active model, and give agents, crews and automations their own.
Open Models & providers with /providers or from Preferences: "Where models come from: sign in, pick a default model, and choose the active one".
Providers
| Provider | Sign-in |
|---|---|
| Anthropic | API key |
| ChatGPT | Sign in with your ChatGPT account (device sign-in) |
| GitHub Copilot | Sign in with your GitHub account (device sign-in). KodaCode uses your Copilot plan. |
| OpenAI | API key or OAuth |
| API key or OAuth | |
| DeepSeek | API key |
| Qwen | API key |
| OpenRouter | API key or OAuth |
| Ollama Local | Nothing: localhost:11434 |
| Ollama Cloud | API key |
| Custom | Any compatible endpoint, with a key |
Add a provider picks an endpoint from the catalog or takes a URL.
- Test Connection checks the provider.
- Save & Set as Active makes it the active provider.
- Remove provider deletes it.
KodaCode stores keys and sign-ins in your keyring.
The GitHub Copilot model list shows the models that your Copilot account turns on. Some of these models may not run on your plan. A request to such a model fails.
Advanced
| Field | Use |
|---|---|
| Base URL / HTTP Proxy | Send a provider's requests to a proxy or a different address. |
| Context window override | When the provider doesn't report the model's window. |
| Stall timeout | How long KodaCode waits on a silent stream before it treats the stream as dropped. |
| The answer arrives as reasoning | For models that put their answer in the reasoning channel. |
| OAuth Client ID | A client registered for KodaCode, if the provider needs one. When this field is blank, GitHub Copilot uses a built-in client. |
| Copilot integration ID | For GitHub Copilot. The ID must belong to the OAuth client ID. |
| Codex client version | For ChatGPT sign-in. |
| VS Code version, Copilot Chat version | For GitHub Copilot. GitHub refuses versions that it no longer supports. If sign-in or requests start to fail, set newer versions. |
| Request body (JSON) | Extra fields sent with every request. |
Choose a model
/models or ⇧⌘L opens Choose a model. There you can:
- pick a provider
- Search models
- filter to Only models with Tools, Reasoning or Vision
- return to Use the default
The list shows recent choices first. The model picker in the top bar sets the model for the next message. Each project remembers its own model.
A model per job
| Where | Set it in |
|---|---|
| An agent | Its Model field on the Agents screen |
| A review part | The review agent on the Agents screen |
| A crew | Leader, worker, stronger and review models in Crew Workflow. Each workflow has its own worker model. |
| An automation | Its decision model in the automation editor |
Reliability and cost
KodaCode retries a request that fails with a rate limit, overload, server error or dropped connection (3 attempts). If the provider is down, the turn then waits 15 to 60 seconds and tries again, up to 5 times. See General settings.
Anthropic models use a prompt cache, 5 minutes by default. Auto uses the 1-hour cache for sessions that you type into. /metrics shows what each provider cost.