Octarine Logo

Octarine guide

How bring-your-own-key AI works in Octarine

How to connect an AI API key to Octarine, which note context is sent, where credentials are stored, who charges for model usage, and when to use a local model instead.

By Rajat Kulkarni · Reviewed August 17, 2026

Bring your own key, often shortened to BYOK, changes who supplies the AI service. Octarine provides the writing and workspace interface, but your chosen provider supplies the model. You create the provider account, obtain its API credential, accept its terms, and pay any usage charges attached to that account.

AI features in Octarine require an active Pro or trial license. That license enables the integration; it does not turn an external provider’s paid API into an included service.

Connect a provider

Open Settings → AI → API Providers, select a provider, enter the required API key, and activate the connection. Octarine validates API-key providers and shows the available models after a successful connection.

Built-in provider options currently include OpenAI, Anthropic, DeepSeek, Google Gemini, Groq, Mistral AI, Perplexity, Together AI, and xAI. Local Ollama and LM Studio connections appear in the same area but use a server URL instead of a cloud key. The current list and setup flow are maintained in the Configuring AI documentation.

If a service exposes an OpenAI- or Anthropic-compatible API, you can add a custom provider. Choose the matching format, give the connection a name, enter its base URL, and add a key when required. Octarine checks the models endpoint before saving it. A custom endpoint deserves the same scrutiny as a built-in provider: the operator of that endpoint receives whatever you send to it.

What is sent during a request

Octarine does not upload the workspace to create a server-side copy. Search, embeddings, and retrieval remain on the device. For each AI request, it sends the prompt and the context included for that turn to the active provider.

Context can include the current note, selected text, or notes and folders you explicitly add with an @ mention. Ask Octarine can search locally and gather relevant passages before making the model request. This narrows the material, but it does not make a request to a cloud model private from that provider.

Use the smallest context that can answer the question. For a sentence-level rewrite, select the sentence rather than sending an entire journal entry. In Ask Octarine, add sensitive paths to Settings → AI → Ignored when they should not be indexed or offered as context. Those exclusions are stored in a readable .octarineignore file at the workspace root.

Where the key is stored

API keys, provider URLs, selected models, and workspace AI rules are stored in Octarine’s device-specific settings file, .store.dat, outside the workspace. They are not copied by Git Sync or ordinary cloud-folder sync of the workspace. That separation helps prevent a credential from being committed alongside notes, but it also means a new computer needs its own provider setup.

Protect the operating-system account and disk that hold the settings file. Do not paste a key into a note, template, Git repository, screenshot, or support message. If a key is exposed, revoke it in the provider’s console and create a new one.

Billing and limits belong to the provider account

The provider sees requests under your credential and applies its current usage rates, quotas, model access, and regional availability. Set a budget or usage alert in the provider console when one is available. A key with an unlimited spend ceiling is unnecessary for a personal notes workflow.

Octarine can show models returned by the connection, but it does not control a provider changing a model name, retiring a model, rejecting a request, or altering a price. If a model disappears from the selector, verify the provider account and its current model access before treating it as a note-storage problem.

Account connections are related, but not the same

Octarine also has an AI → Accounts area for supported assistant accounts. For example, a Codex connection uses the local Codex CLI session instead of asking you to paste an OpenAI API key. That is an account-based connection, not BYOK in the narrow API-key sense. Authentication, eligible plans, and available models follow the connected account.

Use API Providers when you specifically want API-key billing and control. Use Accounts when you want a supported local account session. Confirm the boundary in Settings before sending private context.

When a local model is the better option

Use Ollama, LM Studio, or a compatible local endpoint when note context should not be sent to a cloud provider. The request stays on your computer, but you take on model installation, disk usage, memory requirements, and performance. Local models may also lack provider-hosted tools such as web search.

A practical setup can use both: a local model for sensitive summaries and a cloud provider for non-sensitive tasks that need a stronger model. Octarine’s model selector lets you choose per conversation. Before using either path on real notes, test with harmless content, inspect the selected provider and context, set a provider budget, and document which workspace folders should never be sent.