Your own OpenAI key
On your own server, one OpenAI API key is enough for the assistant on the side, the AI helper, the morning summaries and drawing pictures. You pay OpenAI directly and your data goes only to OpenAI, never to us. This page takes you from the key to the first drawn picture in about ten minutes.
On a hosted instance (killBottleneck Cloud) use Own model in the cloud instead: the OpenAI address and the key go into Organization settings there, and drawing stays on our service.
What you need
- An API key from platform.openai.com → API keys (it starts with
sk-). A ChatGPT subscription is not enough, the API is billed separately. - Credit or a payment method on the API account. Set a monthly limit there too (Settings → Limits), so a mistake can't run up a big bill.
- For drawing: the account must have access to image models. OpenAI may ask you to verify the organization first (Settings → Organization). Until then it rejects drawing, and the picture card shows the reason.
The assistant
In Organization settings → AI features pick the OpenAI-compatible card and fill in:
| Field | Value |
|---|---|
| API address | https://api.openai.com/v1 |
| Model | gpt-5.6-luna (measured, see below) or gpt-4o-mini |
| Token | your sk-… key |
Then click Test connection and Save. The same thing in .env (handy for automated installs):
KB_AI_PROVIDER=openai
KB_AI_URL=https://api.openai.com/v1
KB_AI_TOKEN=sk-...
KB_AI_MODEL=gpt-5.6-lunaThe same key also does voice notes (OpenAI transcribes them) and, with The assistant reads images switched on, reading photos in the assistant. All the options are on the configuration page.
Reasoning models work without extra settings
gpt-5.6-luna and other gpt-5* models are reasoning models. OpenAI refuses to combine their thinking settings with tool calling, and the assistant calls tools all the time (creating steps, events, rules). Up to v1.0.0 that ended with error 400, and you had to add KB_CHAT_OPENAI_EXTRA={"reasoning_effort":"none"} by hand.
Since v1.0.1 the app handles it itself. When OpenAI refuses the setting, it retries once with thinking turned off, and if needed once more without the setting. Your instance remembers what worked for each model, so later turns go through on the first call. If you already have KB_CHAT_OPENAI_EXTRA in .env, you can keep it or delete it. It still takes precedence.
What we measured with gpt-5.6-luna
On a fresh account, in Czech, with the assistant answering and confirming like a user would:
- 20 everyday tasks (a project with deadlines, a meeting, an e-mail, a presentation, a page with a picture, moving a deadline, a rule for overdue steps…): about 8 seconds per task on average. Writing a whole page takes longer, about 90 s.
- "Remind me tomorrow at 8 to call the bank": 10 of 10 runs created a calendar event with a reminder. (Before v1.0.1 it was 3 of 10. The model claimed it couldn't do reminders and only saved an idea.)
- "Every Monday at 9 I have a team meeting, put it in the calendar": 10 of 10 runs created a series of Mondays. (Before v1.0.1 it was 2 of 10, mostly a single meeting.) When you don't say until when, the assistant usually asks first (Next 4 weeks, Until the end of the year).
gpt-4o-mini remains a cheaper choice; it passed every basic AI feature in our August measurement. We have not run the 20 assistant tasks above on it.
Drawing pictures
Drawing has its own variables, so you can use a different key (or a different service) for it:
KB_IMAGE_PROVIDER=openai
KB_IMAGE_TOKEN=sk-... # may be the same key as KB_AI_TOKEN
KB_IMAGE_MODEL=gpt-image-1-mini # default since v1.0.1
KB_IMAGE_TEXT=1 # optional: allow pictures with textRestart the app (docker compose up -d). Then the Pictures tab appears in Office and the assistant starts drawing on request.
What we measured with gpt-image-1-mini:
- One picture takes 15–40 seconds (a local image model is faster, OpenAI is slower).
- Text in the picture came out readable, including Czech accents ("PORADA V PONDĚLÍ 9:00", "Kafe U Lišky").
- OpenAI draws only a few sizes. An e-mail banner (1200 × 400) therefore comes out as the nearest wide size, 1536 × 1024.
Why gpt-image-1-mini and not gpt-image-1
Up to v1.0.0 the default was gpt-image-1. Many new API keys can no longer use it, and drawing failed straight away with "The image service rejected the request (invalid_value)". Since v1.0.1 the default is gpt-image-1-mini, which worked on our test key that no longer had gpt-image-1. If you want a different model, set it in KB_IMAGE_MODEL.
More about pictures (styles, the viewer, pictures in pages and in the morning e-mail) is in Pictures from the assistant.
When something fails
| What you see | What to do |
|---|---|
| Test connection fails with 401 | Wrong or deleted key. Create a new one and paste it with no spaces. |
| 429 / insufficient_quota | The API account has no credit, or you hit the monthly limit. Top up at platform.openai.com. |
The assistant answers, but error 400 mentions reasoning_effort | You're running v1.0.0 or older. Update (Updating) or add KB_CHAT_OPENAI_EXTRA={"reasoning_effort":"none"}. |
| The picture card says rejected the request | Your key doesn't have access to the model. Check KB_IMAGE_MODEL, or verify the organization at OpenAI. |
| No Pictures tab | KB_IMAGE_TOKEN is missing, or the app wasn't restarted after editing .env. |
Where your data goes
Everything you send the assistant goes to OpenAI: your messages, excerpts from maps, the material for pages and pictures. It never passes through us. What OpenAI does with it is set by its API terms. Your key stays on your server and is never sent back to the browser.
Next
- Configuration → AI: other providers, spending caps, a separate model for summaries
- Environment variables: every
KB_AI_*andKB_IMAGE_*option - The AI assistant on the side: what the assistant can do

