AI assistant

Ask a question in the launcher and read the answer without leaving it. Type ai and a question, press Enter, and the answer appears as a result you can copy, paste or read in full.

Off by default, and your questions leave this computer

The assistant is off until you turn it on (Settings → AI assistant). With Ollama the question stays on your computer. With OpenAI-compatible or Anthropic, the question is sent over the internet to that service when you press Enter, and the service's own privacy terms apply to it. Exactly what is sent is listed below.

Set it up

1. Open Settings → AI assistant.
2. Choose a provider and switch Use the AI assistant on.
3. For OpenAI or Anthropic, paste an API key and press Save key. For Ollama there is no key.
4. Press Test connection to check the address, the key and the model, then Save.

Provider	Talks to	Key	Default model
Ollama (the default)	http://localhost:11434, a model running on your computer	none	llama3.2
OpenAI-compatible	https://api.openai.com/v1 (ChatGPT / OpenAI), or the base URL you set	OPENAI_API_KEY or a saved key	gpt-4o-mini
Anthropic	https://api.anthropic.com	ANTHROPIC_API_KEY or a saved key	claude-haiku-4-5

The default models are only a starting point: providers retire models, so type the name you want into Model.

Ollama needs the Ollama app running and the model downloaded once (ollama pull llama3.2). Test connection says if the model is missing.

OpenAI-compatible works with any server that copies OpenAI's chat API: set the Base URL, for example http://localhost:1234/v1 for LM Studio, https://openrouter.ai/api/v1 for OpenRouter or https://api.groq.com/openai/v1 for Groq. A server on your own computer needs no key. Azure OpenAI uses a different sign-in header and is not supported.

Using it

You type	What you see	Enter
ai	a hint saying where a question will go	nothing
ai why is the sky blue	Ask "why is the sky blue", with the model and host it will be sent to	sends the question
(while it waits)	Asking gpt-4o-mini…	nothing; the action panel (Ctrl+K) has Cancel the question
(the answer)	the answer, then Open full answer for a long one	copies the answer
(a failure)	a row that says what went wrong	tries again

On the answer: Shift+Enter pastes it into the app you were using (where pasting works), Ctrl+T or Open full answer reads it all in the text view, and the action panel (Ctrl+K) has Ask again. The launcher stays open while the question is on its way, and the answer replaces the Asking… row.

Typing never sends anything: only Enter on the Ask row does. There is no streaming; the answer appears when it is complete. Each question is separate: the assistant does not remember earlier ones.

Ask AI about selection

Select text in any app and press the Universal Actions key. Ask AI about selection puts ai Explain this: <your selection> in the search box so you can edit it. Nothing is sent until you press Enter there. It is offered only for text, only when the assistant is on, and a selection longer than 2,000 characters is cut at that length.

What is sent, and to whom

When you press Enter on an Ask row, Sevak makes one HTTPS request to the base URL's host containing:

• your question (or the text you chose with Ask AI about selection, after you pressed Enter),
• the system prompt from Settings,
• the model name and the maximum answer length,
• your API key, in the header the provider expects (never in the address or the body).

Nothing else goes with it: not your clipboard, your files, your search history, the apps you use, or an identifier. The only extra header is User-Agent: Sevak/<version> (ai).

Provider	Host that receives it	Request
Ollama	localhost:11434 (or your base URL)	POST /api/chat
OpenAI-compatible	api.openai.com (or your base URL)	POST /chat/completions
Anthropic	api.anthropic.com (or your base URL)	POST /v1/messages

Test connection makes a different request: GET of the provider's model list (/models, /v1/models or /api/tags), with the key if there is one. It sends no question and runs only when you press the button.

Sevak makes no AI request in any other situation: not at startup, not while you type, not when you open the launcher, and not in the background. There is no telemetry, no retry loop and no caching of what you ask. See Privacy.

What is kept

The question and the answer live in memory until you ask something else, reload the settings or quit Sevak. They are not written to disk, not logged, and not added to the search history or usage statistics. (The log records only that a question was asked, the provider and how long it took.)

What the provider keeps is up to the provider: see the terms of OpenAI, Anthropic or the service behind your base URL.

API keys

• A key is never written to config.toml, a settings export or the diagnostics report, and never to the log.
• Keys you save in Settings are kept in ai-keys.json in the data folder:
  • on Windows, encrypted with DPAPI for your user account (only your Windows account on this computer can read it back);
  • on macOS and Linux, in a file only your user account can read (mode 0600). It is not encrypted there, like the clipboard history; Sevak does not use the macOS Keychain or the Linux Secret Service.
• If nothing is saved, the OPENAI_API_KEY or ANTHROPIC_API_KEY environment variable is used. A key saved in Settings takes precedence.
• The key field is write-only: once saved, it is never shown again. Remove deletes it.
• A key is never sent over plain http:// to another computer: Sevak refuses and tells you to use https://. Redirects are never followed, so a key cannot be passed on to another address.

Safety

• The answer is untrusted text. It is shown and copied as plain text. Nothing in it is run, opened or interpreted, and Sevak never creates a result that launches, opens or executes anything from it. Control characters and text-direction overrides are removed first.
• Everything is capped. Questions: 8,000 characters. The system prompt: 4,000 characters. A request: 64 KiB. A reply from the server: 1 MiB (a larger one is discarded). An answer: 20,000 characters. The wait: the Timeout setting (5 to 300 seconds).
• Offline is fine. Without a connection, or with Ollama not running, you get a clear message, and the rest of Sevak is unaffected.
• Questions you cancel stop being shown at once. The request itself runs until its timeout, but its reply is thrown away.

Troubleshooting

Message	What to do
Ollama is not running	Start the Ollama app (or ollama serve), or fix the base URL.
That model was not found	For Ollama run ollama pull <model>; otherwise check the model name.
No API key for …	Paste a key under Settings → AI assistant, or set OPENAI_API_KEY / ANTHROPIC_API_KEY and restart Sevak.
The AI service did not accept the API key	The key is wrong, revoked or lacks access. Save it again.
The AI service is busy or over its limit	Wait, or check your credit and rate limits with the provider.
The AI service took too long	Raise Timeout; a local model may still be loading.
The answer was too large	Lower Longest answer; the server sent more than 1 MiB.
The model used its whole token budget before answering	Raise Longest answer. This happens with reasoning models.
The stored API key could not be read	It was saved by another Windows user or computer. Enter it again.
The AI assistant is off	Switch it on in Settings → AI assistant (and make sure AI assistant is on under Plugins).

Options

All of these are in Settings → AI assistant; the [ai] section of config.toml has the same keys. The API key is the one thing that is not in the file.

Setting	Default	What it does
Use the AI assistant	off	Nothing is sent unless this is on
Keyword	ai	What you type before the question
Provider	Ollama	ollama, openai or anthropic
Model	provider's default	Which model answers
Base URL	provider's default	Where requests go
System prompt	a short "be concise" instruction	Sent with every question
Longest answer	512 tokens	16 to 8192
Timeout	60 seconds	5 to 300

Turn the plugin off completely by adding "ai" to [plugins] disabled.

Not included

Streaming answers, conversation memory, images and files, tool use, and Azure OpenAI's sign-in are not supported. The assistant answers one question at a time and never acts on your computer.