Set up AI

Oleafly’s AI is bring-your-own. Connect a provider you already use or a model running on your machine. Keys stay encrypted on disk. The Rust backend resolves the selected provider, makes the request, and streams the result to the interface. There is no Oleafly cloud service in the middle and no markup on your tokens.

Open Settings → AI Assistant, or use Connect a provider from the chat panel. AI settings are split into three tabs: Providers and keys, Instructions, and Personas.

Built-in providers

Provider	Notes
OpenAI	GPT-4o family, GPT-4.1 family, o-series when available
Anthropic	Claude models (Sonnet, Haiku, and current lineups)
Google Gemini	Gemini models from Google AI
Perplexity	Perplexity chat models
Z.AI (GLM)	GLM coding and chat models
Groq	Fast Llama and related models
OpenRouter	Routed access to many hosted models
DeepSeek	Chat and reasoner (R1) models
Mistral	Large, Codestral, Small, and related
xAI (Grok)	Grok models
Ollama (local)	Whatever you have pulled on this machine

When you save a key, Oleafly validates it and loads the provider’s current model catalog. Each card has a Get key link to the provider’s signup page. Some reasoning models can return a separate thinking stream. See Chat and tools.

Connecting a provider

1. Open Providers and keys.
2. Expand a provider, paste the API key, and save.
3. Oleafly checks the key and refreshes the model list. A green Connected badge appears.
4. Under Instructions, set the default chat model if you want a standing preference. The chat panel can still switch models for a single conversation without changing that default.

You can connect several providers at once. Every connected provider shows the same green Connected badge. You do not need a second click to make a saved provider usable. Pick models from the chat header or the default-model control. Use the trash control on a card to remove a key.

You can also toggle individual models on or off per provider, add a model by id, delete with confirmation, or restore that provider’s built-in list.

Custom providers (OpenAI-compatible)

If you run a local server (LM Studio, a self-hosted gateway, another Ollama-compatible OpenAI API) or any other OpenAI-style base URL:

1. Add a custom provider.
2. Enter the base URL and, if needed, an API key.
3. Save. Oleafly validates the endpoint the same way as built-in providers when possible.

Use this when your endpoint is not one of the built-in brands. For the standard local Ollama install, the dedicated Ollama card is usually simpler.

Going local with Ollama

1. Install Ollama and pull a model, for example ollama pull llama3.2.
2. On the Ollama card, click Check for Ollama. Oleafly finds the local server and lists installed models.
3. Pick a model. No cloud key is involved.

If Ollama is not running, the card shows what to run and a Re-check button. A custom host for Ollama on another machine lives under advanced host settings on that card.

Instructions

The Instructions tab holds standing preferences for chat and inline edits, for example: “Prefer British English. Keep answers short. Use enumitem for lists.”

Those preferences are sandboxed. They steer tone and style. They cannot override tools, approval rules, or safety behavior.

The tools catalog and Allow PDF page capture toggle also live here. PDF capture lets the assistant rasterize pages for vision checks. Leave it off if you do not want that.

Personas

Personas are named instruction presets with a color. Create them under Personas, then switch them from the chat Prompts menu. A create shortcut can deep-link into Settings when you want a new persona mid-work.

Use personas for roles you reuse, such as “strict reviewer”, “clearer prose”, or “notation cop”. They work alongside your global instructions, and the app shows which one is active.

Where your keys live

• Keys sit in AES-256-GCM encrypted, owner-restricted files under ~/.oleafly/. Provider credentials are resolved in Rust and are not sent to the webview. They leave the machine only in requests to the provider they belong to. Encryption prevents plain-text storage, but it cannot protect an account that is already compromised.
• Requests go from the Rust backend to the provider over HTTPS, or to localhost for Ollama and many custom endpoints.
• Chat history stores attachment names and types, not the attachment bytes.

What the assistant can do once connected

The tools list (read, edit with approval diffs, compile, log, literature search, and more) is documented on Chat & tools. Every file-changing action still waits for your approval.