Providers & BYOK

Cordy for Mobile never proxies your traffic. You bring an API key, Cordy talks directly to that provider from your device, and your key is stored in secure device storage — never in the conversation database or logs. This page covers every provider option and how the model catalog and key storage work.

The BYOK principle

• You own the account. Keys come from the provider's console; you pay the provider directly.
• One key per provider. Each provider stores its own independent key. Switching providers switches keys.
• Direct requests. Prompts go to the selected endpoint using that provider's authentication header. Built-in and public custom endpoints require HTTPS; allowed local endpoints may use HTTP. Nothing is routed through a Cordy service.

Built-in providers

Cordy ships with five built-in provider presets. OpenRouter is selected by default.

Provider	API family	Base URL	Notes
OpenRouter (default)	OpenAI-compatible	https://openrouter.ai/api/v1	One key, 100+ models; catalog includes live pricing
OpenAI	OpenAI-compatible	https://api.openai.com/v1	First-party OpenAI models
Anthropic	Anthropic (native)	https://api.anthropic.com	Native API with extended thinking
Vercel AI Gateway	OpenAI-compatible	https://ai-gateway.vercel.sh/v1	Routes to models behind your Vercel gateway
Volcengine Ark (火山方舟)	OpenAI-compatible	https://ark.cn-beijing.volces.com/api/v3	Volcengine's model service

OpenRouter

The default and the easiest first key. OpenRouter is an OpenAI-compatible aggregator: a single key reaches models from many vendors. Its /models response includes pricing, so Cordy shows real per-model costs. Model IDs use a vendor/model form, for example anthropic/claude-sonnet-4-5.

OpenAI

First-party access to OpenAI's models via api.openai.com/v1. Model IDs are plain names (for example gpt-4o). Pricing for OpenAI models is resolved from the bundled LiteLLM snapshot when the /models response doesn't include it.

Anthropic (native)

Anthropic uses Cordy's native Anthropic client rather than the OpenAI-compatible path. This enables extended thinking through the reasoning-level control. Because it is the native API rather than an OpenAI-style /models endpoint, Anthropic's bootstrap model entry is built in.

Vercel AI Gateway

An OpenAI-compatible gateway; point Cordy at it with a Vercel AI Gateway key and reach whatever models your gateway exposes.

Volcengine Ark (火山方舟)

Volcengine's OpenAI-compatible model service (ark.cn-beijing.volces.com/api/v3). Use an Ark API key.

Enabling and disabling providers

Under Settings → Availability, you can toggle any provider on or off. A disabled provider is hidden from the chat model picker until you re-enable it. This is handy for keeping the picker focused on the providers you actually use.

Custom OpenAI-compatible endpoints

Beyond the built-ins, you can add your own OpenAI-compatible endpoint — a self-hosted model server, an in-house gateway, or another vendor that speaks the OpenAI API.

To add one: Settings → Custom provider → enter a Name and a Base URL (the field hints "Base URL (ends with /v1)") → Add provider. It then behaves like any built-in: save a key, load its catalog, and select its models. Custom providers carry a Custom badge and are the only providers you can delete.

Base-URL rules

The URL is parsed and validated before the endpoint is saved, so a malformed or plaintext URL can't quietly send your key to the wrong host:

• It must be a valid http(s) URL with a host.
• HTTPS is required for public hosts — your key is never sent in the clear.
• Plain http:// is allowed only for loopback / private-LAN development hosts.
• Embedded credentials (user:secret@host) are rejected.
• A trailing slash is trimmed; the normalized URL is what's stored.

Allowed plaintext (http://) hosts, for local development:

Host pattern	Example
localhost	http://localhost:11434/v1
127.0.0.1 / 0.0.0.0	http://127.0.0.1:1234/v1
10.x.x.x	http://10.0.0.5:8000/v1
192.168.x.x	http://192.168.1.20:8080/v1
172.16.x.x–172.31.x.x	http://172.16.0.10:8000/v1
*.local	http://my-box.local:8000/v1

Any other host must use https://. This lets you point Cordy at a model server on your own machine or LAN while keeping every cloud endpoint encrypted.

The model catalog

When a provider has a saved key, Cordy fetches and caches its model list:

• Fetch. For OpenAI-compatible providers, Cordy calls GET {baseUrl}/models with your key and maps each model's context window, capabilities, and pricing. Anthropic (native) uses its built-in entry instead.
• Capabilities. Each model is tagged from the response (and a bundled LiteLLM snapshot when the response is sparse) with Vision, Thinking / reasoning, Tool use, and Long context (context window ≥ 100k). These tags drive the UI — for example, image send is only allowed on Vision models, and the reasoning-level control only appears on Thinking models. A Tool-use tag does not enable execution in the app.
• Caching. The fetched catalog is cached on device, so the list is instant next launch and available offline. Use Refresh models to re-fetch.
• No key yet. Before you save a key, Cordy shows whatever is cached (or a small bootstrap list) without hitting the network.

Pricing

Prices are normalized to USD per 1,000,000 tokens and resolved in this order:

1. Provider-native — pricing carried in the provider's /models response (OpenRouter provides this).
2. LiteLLM snapshot — a bundled price table, looked up by normalized model-ID variants (tolerating vendor prefixes and OpenRouter suffixes).
3. Unknown — the model shows — / "Price unknown".

Cordy displays these as "In $… / Out $… per 1M" in the model list, and records cost per response on the usage page. Treat in-app costs as estimates and reconcile against your provider's own billing.

Model ID formats

• OpenRouter: vendor/model (for example anthropic/claude-sonnet-4-5).
• OpenAI / Vercel / Ark / custom: the provider's own model names.
• Anthropic native: Anthropic model IDs (for example claude-sonnet-4-5-20251022).

How your keys are stored

API keys are held only in platform secure storage, scoped per provider:

• Android — a Keystore-held AES-256-GCM key encrypts each value, stored as Base64(IV ‖ ciphertext ‖ GCM tag) in a private preferences file. Values written by the older EncryptedSharedPreferences are transparently migrated to the new scheme on first read, so no key is lost across an app upgrade.
• iOS source — uses Keychain with AfterFirstUnlockThisDeviceOnly accessibility, intended to keep the value device-bound and out of iCloud Keychain and backups. The repository records no completed iOS binary link, execution, or smoke, so this has no runtime verification.

Keys are never written to the conversation database or application logs. They leave the device only as authentication material on requests to the provider you chose. Saving a blank key is treated as deleting it. For the complete data-handling story, see Privacy & security.

Not yet available

• Tool / function calling is modeled in the message format (the app can render tool-use and tool-result blocks) but is not executed in this release.
• Automatic multi-provider gateway / fallback routing is future work.

Today, each conversation talks to exactly one provider and one model.