AI providers

Dieser Inhalt ist noch nicht in deiner Sprache verfügbar.

Kept works without AI. Connecting a provider adds three things: naming things from photos and reading receipts, labels and meter readings; the assistant, which answers questions about your inventory; and, with an embeddings model, search by meaning. You bring your own key; Kept has no AI service of its own.

Providers

Provider	Photos and receipts	Assistant	Search by meaning
Groq (recommended)	yes	yes	no embeddings model
OpenAI	yes	yes	yes
Anthropic	yes	yes	no embeddings model
Google (Gemini)	yes	yes	yes
OpenRouter	yes	yes	no embeddings model
OpenAI-compatible (Ollama, LM Studio, …)	the models you name	the models you name	if the server offers one

Groq with qwen/qwen3.8-27b is the recommendation: the cheapest reliable result Kept measured, about USD 0.004 a receipt (on synthetic test images, 2026-09-26). Pasting a Groq key selects it, and Why? on the page gives the figures. Kept suggests a model for each provider and task; you can pick any other from the provider’s list under Advanced.

Where the key goes

There are three places for a key, and no per-location key:

Key	Set in	Pays for
The account’s key	Settings → AI, by the account’s owner	Every location that account owns
Your personal key	Settings → AI → Personal AI key	Your Personal location, your private assistant threads, and questions that span owners
This server’s key	Admin → AI, by an instance admin	Anyone whose account has no key of its own

Work in a location (naming a photo, an assistant answer about it) is paid by the location’s owner, so members of Alfred’s “Family house” use Alfred’s key, not their own. Members and viewers see a read-only line saying who pays.

Connecting: paste the key in Paste your key. Kept recognises the provider from the key where it can (otherwise choose it under Advanced), then Save and test checks that the model reads photos and gives structured answers, and says what the test cost. Keys are write-only: Kept never shows a saved key again, and Replace key swaps it. Before the first key on a server, the instance admin has to save the recovery kit: the key is a secret Kept must be able to recover from a backup.

Advanced holds the provider and base URL, the model for each task (Models, loaded from the provider’s own list), Reasoning effort (low by default), the limits and the prices.

What uses AI, and what it costs

AI settings open with What uses AI in Kept: a row per action (a captured photo, a receipt, a label, a reading, an assistant question, a search, the test), with how many calls it makes, the tokens it takes, and the cost from the price table, from your last 30 days’ use once there is some. Under This month, a bar per limit shows the spend so far.

Capture (Thing, Receipt, Label, Reading modes): one call per photo. Names, brands, models and types are filled in; prices, dates, serials and readings wait in the Inbox as suggested values for you to accept. Each AI-filled draft carries a line such as “Read by qwen/qwen3.8-27b (Groq) · 2,502 tokens”, with its cost and who paid.
Several things in one photo (Thing mode; Admin → Settings, off by default): the same one call names up to 12 things in the photo. The first fills in the capture’s own draft and keeps the photo; each other one becomes a draft beside it, in the same place and batch, with its own picture cut from the photo, to confirm, merge or discard in the Inbox. How well the pictures are cut depends on the provider’s boxes, so check a few drafts after turning it on.
The assistant answers from your inventory within your role, and each tool it uses adds a call. Changes it proposes show as a confirmation card that you approve; for viewers it is read-only.
Search words after an import: an import can offer Add search words with AI?, with the token estimate first (import).

AI usage pages list every call: My AI usage for everyone, a location’s usage for its admins and owner, the account’s for its owner, and the instance’s under Admin → AI. Each shows totals, tokens by day stacked by task, breakdowns by task, person and location where they apply, and the call list with Export CSV. A call records its model, tokens, outcome and cost, never the prompt, the photo, the answer or the key.

Prices come from the price table, which instance admins edit in Admin → AI (with Fill from … listing for providers that publish prices). The recommended Groq model’s price is filled in for you. A call with no price is counted in tokens and shown as cost unknown.

Limits and pausing

Nothing is capped by money unless someone chooses to: after the first key Kept offers Set a monthly limit? with a suggested amount, and No limit. Under Advanced, Limits holds:

Money a month (in a currency the server has turned on) and Tokens a month, for the account; a location’s own cap inside the account’s, set by the owner; and a limit per person.
Photo and receipt reading has a token budget of 60,000 tokens a minute, 2,000,000 a day and 20,000,000 a month per paying account unless changed.
Pause AI stops all calls by hand until you resume.

When a limit is reached, AI pauses for that scope only, with a banner such as “AI paused until 1 Oct · Home’s monthly cap reached”. Captures still save; naming waits, and resumes when AI does. Whoever set the cap gets Resume now, to raise it, remove it or keep paused. A provider’s own rate limit is not a pause: progress lines say “Waiting for Groq · about 20 s” and carry on.

One assistant request is kept under the provider’s per-minute limit once the provider has named it (Groq’s free tier: 8,000 tokens a minute). Kept first asks for a shorter answer, then leaves out the oldest tool results from what it sends; the thread keeps them. A question that still can’t fit fails with “too large”: a model or plan with a higher limit answers it.

What is sent to the provider

Kept’s own summary, shown beside each provider:

Only a copy of the photo with location and camera data removed, never the original. The fixed instructions and your location’s languages. For the assistant, your question and the inventory it looks up, within your role. Never secret fields.

Under it, Kept summarises what the provider says it does with API data (training, retention) and links its policy. A server you run yourself keeps what you let it keep.

Search and embeddings

Keyword search always works, including the search words AI adds to things in each of the location’s languages. Search by meaning needs an embeddings model, which comes from one of three places, set by the server’s admin with KEPT_EMBEDDINGS (configuration):

provider (the default): the location’s provider, when it has an embeddings model (OpenAI, Google, or an OpenAI-compatible server). Groq, Anthropic and OpenRouter don’t, so with those, search stays keyword only and says so.
local: a small model inside Kept, with no key and no per-call cost. Kept refuses this setting unless the local model’s runtime is installed.
off: keyword search only.

Local models with Ollama

To keep everything on your own hardware, run a model with Ollama and add it as an OpenAI-compatible provider with its base URL; Kept fills in nothing for this kind, so you name the models. An address on your private network has to be allowed by an instance admin in Admin → Settings first. Setup, memory and the Compose profile: Ollama.

Turning AI off

For one location: Location settings → What to track, under AI in this location, switch off AI capture and the assistant. Capture still works; things wait in the Inbox for a name.
For a while: Pause AI, under Advanced.
Not connecting at all: with no key at any level nothing is sent, and the checklist’s Connect an AI provider step can be ignored.

The AI page has no button to delete a saved key; replace it, pause, or turn AI off per location.