Providers & models

Connect your own provider keys (11 built-in providers or any OpenAI-compatible endpoint) and route a model to each feature.

Cordy is bring-your-own-key. You connect an API key for a provider you already use, and Cordy calls that provider directly from your browser — there is no Cordy proxy or account in between. You can connect several providers at once and assign a different model to each feature.

Built-in providers

Cordy ships with 11 cloud providers built in. Each has a pre-filled default Base URL (which you can override) and a masked API-key field. Create the key in the provider's own console; its documentation explains how.

ProviderDocumentationNotes
OpenRouteropenrouter.ai/docsOne key for models from many vendors
OpenAIplatform.openai.com/docsGPT models
Anthropicdocs.anthropic.comClaude models
Googleai.google.dev/docsGemini models (Google AI Studio)
xAIdocs.x.aiGrok models
Vercel AI Gatewayvercel.com/docs/ai-gatewayOne key in front of many providers; Base URL is fixed
Volcengine Arkvolcengine.com/docs/82379China-region models; the provider card is labeled 火山方舟
DeepSeekapi-docs.deepseek.comAdded in 2.5.5
Qwen (Alibaba Cloud Model Studio)help.aliyun.com/zh/model-studioAdded in 2.5.5; the default Base URL is the mainland China endpoint (dashscope.aliyuncs.com)
Groqconsole.groq.com/docsAdded in 2.5.5
Mistral AIdocs.mistral.aiAdded in 2.5.5

You don't need all of them — one is enough to start. OpenRouter is a good first choice because a single key unlocks many models.

Custom OpenAI-compatible providers

Beyond the built-ins, you can add your own OpenAI-compatible endpoint from the Custom Providers section (in the dashboard). Give it a display name, a Base URL, and an API key, then type the model IDs you want to use (one per line). The manual model list decides which models are available. This covers self-hosted gateways and any service that speaks the OpenAI API.

Check model catalog asks the endpoint for its model list, but only as a diagnostic. If the check fails, Cordy says so and the models you entered manually keep working (since 2.5.5).

For local runtimes like Ollama or LM Studio over loopback, use the dedicated Local-Native Runtimes setup instead — see Local & on-device AI.

Adding a key

API keys are managed in the dashboard's provider directory and editor.

  1. Open the dashboard from the side-panel header and go to Settings → AI Config → Provider Configuration.
  2. Search or browse the provider directory, open a provider, paste your key into API Key, and optionally adjust its Base URL. The key is saved when you leave the field.
  3. Click Detect to verify the connection. The status badge updates to Available, Unavailable, Checking, or Not configured, and a successful check refreshes the provider's model list.
  4. Repeat for any other providers you want.

If a provider shows as configured but its key later turns up missing or unreadable (for example, it was cleared with your browsing data), Cordy shows an inline warning asking you to re-enter it.

Per-feature model routing

Different jobs can use different models. Under Settings → AI Config → Feature Models, Cordy gives supported features their own slots:

  • Chat — the main conversation model.
  • Translation — used by the floating menu, command palette, and translation actions.
  • Explanation — used by "explain."
  • Summarization — used by "summarize."
  • Read aloud (TTS) — the text-to-speech engine and voice (see Read aloud).

For each text slot you choose a provider and model. The Chat slot also has a context limit — Standard (200K) (the smaller of the model's window and 200K, the default) or Auto max (the model's full window). A slot can point at a cloud provider, a custom endpoint, or a local runtime.

Model limits

Since 2.5.7, Cordy reads each model's context size and output limit from, in order: the provider's own model list (OpenRouter, Groq, Google, Mistral AI, and Volcengine Ark only; values other providers report are ignored), the public Vercel AI Gateway model catalog, and the public LiteLLM model catalog. When none of them knows the model, the limit is treated as unknown, and each feature applies its own fallback; for example, translation then assumes a 32,768-token context window and 8,192 output tokens. The catalogs are cached locally and refreshed about once a day. The downloads carry no keys or personal data; the privacy policy lists them.

Switching models mid-chat

You don't have to open Settings to change models. The composer shows the current model — click it to search and switch for the current conversation. The model selector is searchable (Esc clears the search), and shows just the model name without the vendor prefix. The compact routing card in the side panel links to the full dashboard settings when you need to add a key.

Local & on-device models

You can route supported slots to a model that runs on your machine — Chrome's built-in Gemini Nano, in-browser Gemma via WebGPU, or an OpenAI-compatible runtime such as Ollama, LM Studio, llama.cpp, vLLM, LocalAI, or Unsloth. Translation does not offer local-native endpoints; its local choice uses the browser-local engine. Setup and requirements are in Local & on-device AI.

Key security

  • Keys are encrypted before they are written to Cordy's IndexedDB secure store. The encryption key is non-extractable; keys are not kept as plain settings.
  • Requests go directly to the provider you configured. Cordy runs no backend and never sees your traffic.
  • Provider-key deletion is a separate Data Management action. Keys are not included in data backups, so you re-enter them after restoring a backup.

See Privacy & permissions for the full data model.

Troubleshooting

  • "API Key Required" / no models listed. The provider isn't configured, or Detect hasn't succeeded. Re-check the key and Base URL.
  • "No models available" after adding a key. Click Detect to refresh the catalog; for a custom endpoint, make sure you've typed model IDs manually.
  • "Model catalog check failed; manual models still work." A custom endpoint didn't return a model list. Your manual models are unaffected.
  • A model rejects a parameter. Some models don't support every option; try another model or provider.
  • Key "missing" warning. Your stored key was cleared or corrupted — re-enter it in Provider Configuration.