Provider matrix
Connection types and capability limits for each Cordy product.
Choose a connection type first, then verify that the selected model supports the operation you need. A provider name alone does not establish compatibility with every feature.
| Connection type | Chrome | Desktop | Mobile | Gateway |
|---|---|---|---|---|
| Cloud model APIs | Built-in provider choices and custom connections | Configured by each external CLI agent | OpenAI-compatible and Anthropic families | Operator-configured providers, channels, and public models |
| Custom OpenAI-compatible endpoint | Supported through provider configuration | Only if the chosen CLI supports it | Supported through the OpenAI-compatible family | Runtime support depends on channel configuration; the current Admin endpoint field has a known limitation |
| Local model server | Local-native connections; permission and runtime checks apply | Depends on the CLI | Reachable endpoint using supported protocol and network policy | An upstream service reachable from the gateway |
| Browser/on-device model | Chrome built-in AI and WebGPU models, subject to browser/device support | No shared browser model runtime | No bundled on-device inference runtime | No bundled on-device inference runtime |
| Cordy Gateway as endpoint | Compatible text requests with a gateway API key | Requires the CLI to support the gateway API | Compatible text requests with a gateway API key | Provides the endpoint |
Product guides
- Chrome models and local AI: provider setup, model capabilities, local-native and browser inference.
- Desktop agents: install/login to a CLI and configure its model access.
- Mobile providers: OpenAI-compatible connections, OpenRouter setup, and direct Anthropic.
- Gateway providers: map upstream provider models to the public IDs returned to clients.
Check these boundaries
- API and model ID. The endpoint must implement the API the client sends, and the model must be available to those credentials.
- Capabilities. Images, tools, reasoning, and generation controls depend on both the client adapter and the model. Gateway currently rejects image/audio inputs; Mobile does not execute tool calls.
- Network location.
localhostrefers to the machine running the client. On a phone it does not point to your desktop model server; inside a container it does not point to your host. Follow the product's network and credential rules. - Data flow. Cloud APIs receive your submitted content. A local server keeps inference local only when that server itself uses local inference.
This table describes supported connection categories, not a claim that every provider or device has been tested. See the individual guides for release and platform verification limits.