The built-in assistant

Ask Exhale in plain language, in the web and in chat, and let it read a scanned statement or a screenshot into the import list.

It is off by default. You bring the model: Claude, OpenAI, or one running on your own machine. Nothing is sent anywhere until an owner turns it on.

If you already use Claude, ChatGPT or another agent, you may not need this: connect it to Exhale over MCP instead (AI agents in the web). The built-in assistant is for asking from inside Exhale and from Telegram.

What is sent, and where

While it is on, each message you type to it goes to the provider you chose, together with:

• a briefing of your household: today’s date, the names of the people you share with, the subscriptions you can see (name, price, cycle, next charge, payment method, category), the next 30 days of charges and what needs attention;
• what Exhale looks up to answer (for example the history of one subscription, including its notes);
• the last few turns of the conversation, for half an hour.

When you import a scanned PDF, a picture or pasted text that Exhale cannot read by itself, that file is sent too. When you press Read dates on a document attached to a subscription, that document is sent, with the subscription’s name. When you import a bank export (CSV) or a PDF with text, Exhale reads it by itself and sends only the merchant descriptions, to tidy the names: no amounts, no dates.

Never sent: passwords, API tokens, the bot token, the model’s own API key, and subscriptions another member marked private.

Nothing goes to the makers of Exhale. There is no telemetry.

Turn it on

Settings → Built-in assistant (owners only). Pick a provider, save, press Test.

Claude

1. Create an API key at console.anthropic.com.
2. Provider: Claude (Anthropic). Paste the key. Leave the model empty to use claude-opus-5-5, or type another, for example claude-sonnet-5-5 or claude-haiku-5-5 (cheaper and faster).
3. Save, then Test.

OpenAI

1. Create an API key at platform.openai.com.
2. Provider: OpenAI-compatible. Paste the key and type the model. Leave the address empty.

The same provider works for OpenRouter (address https://openrouter.ai/api/v1) and any other service that speaks the OpenAI chat completions protocol.

A local model with Ollama

Nothing leaves your network.

1. Install Ollama and pull a model that supports tools, for example ollama pull llama3.1 or ollama pull qwen2.5.
2. Provider: OpenAI-compatible. Address: http://localhost:11434/v1 (from Docker: http://host.docker.internal:11434/v1). Model: the name you pulled. No key.
3. Press Test. If it says the model “did not use the test tool”, pick a model with tool support: questions and changes need it.

LM Studio works the same way with the address it shows (usually http://localhost:1234/v1).

Small local models answer simple questions well and get complex requests wrong more often. Every change still waits for your confirmation, so a wrong one costs a tap on Don’t.

From the environment

For Docker or a service, the same settings as variables. Each one that is set wins over the web and shows there as “Set on the server”.

EXHALE_ASSISTANT_PROVIDER=anthropic        # or openai
EXHALE_ASSISTANT_API_KEY=sk-ant-…
EXHALE_ASSISTANT_MODEL=claude-opus-5-5     # optional for anthropic
EXHALE_ASSISTANT_BASE_URL=                 # e.g. http://ollama:11434/v1

The on/off switch stays in the web.

Ask

In the web: the Ask Exhale box on Overview, or ⌘K / Ctrl+K on any page. In Telegram: just type.

Questions get an answer:

• What do I pay per month?
• What renews this week?
• How much have I paid Netflix in total?
• Which subscriptions are yearly?
• What could I cancel?

Requests get a card that says exactly what would change, with Confirm / Discard (in chat: ✅ Do it / ✖ Don’t). Nothing changes until you confirm, and what you confirmed has Undo:

• Add Netflix, 13.99 on the 6th
• Netflix went up to 15.99
• Move this month’s gym charge to the 20th
• Mark the Spotify charge as paid
• Cancel Disney+ at the end of the period
• Pause the gym

The card is written by Exhale, not by the model: if the model misunderstood, the card shows it. Cancelling here stops the reminders; cancel with the provider too.

When a list says it better, the answer comes with a screen: your subscriptions, the charges of a period, or what you pay per month or year broken down by category, payment method or subscription. The model only chooses what to show; every name, date and amount on it is read by Exhale from your data. Tap a row to open it (in chat, each one is a button). Try:

• Show me all my subscriptions
• What is charged between now and Christmas?
• What do I pay per year, by payment method?

The answer appears as it is written. In Telegram it grows in one message (updated every few seconds, within Telegram’s limits); the cards and screens come when it is finished. Behind a reverse proxy that buffers responses, the web answer may arrive whole instead; it still works.

It answers in the language you write in. The cards, screens, buttons and error messages follow your language: in chat the household’s, in the web your browser’s (or the one you picked).

Import a statement or a screenshot

Settings → Data → Import, or send the file to your Telegram bot.

What	Needs the assistant	How it is read
Bank export of transactions (CSV)	No	Exhale reads it. Two or more months find the charges that repeat
Statement in PDF with text	No	Exhale reads the lines (a date, a description, an amount)
Lines copied from a statement	No	Same as a PDF: paste them in the box
Scanned PDF (no text)	Yes	The model reads it
Screenshot or photo (PNG, JPEG, WebP)	Yes	The model reads it: a subscriptions page in your bank’s app, an App Store list, an email
Other pasted text (an email)	Yes	The model reads it

The list has up to three parts:

• Paid, according to the statement. A payment that matches a charge Exhale expected: the same service, within a week of the date, an amount within 30%. Ticked lines are marked paid. If the bank took another amount, Exhale asks whether it is the new price.
• Expected but not in this statement. Charges due in the period that the statement does not show (only for the cards the statement covers). Worth a look: a failed payment, or a subscription that ended.
• New subscriptions. Charges that repeat on a regular cycle and are not in Exhale yet. Each line says what it is based on (“6 charges every month, last on 2026-11-06”).

Tick what to apply. Nothing is saved before that, and the file itself is never stored.

From Telegram: send the file (as a document or a photo) to your bot. It answers with the same list; tap a number to skip a line, then Accept all or Apply. A list longer than eight lines waits in the web (Settings → Data → Import) for a day, with a link when EXHALE_PUBLIC_URL is set. Slack does not take files yet.

Limits: pictures up to 5 MB, PDFs and other files up to 20 MB, text sent to a model up to 200 KB. A scanned PDF needs a model that accepts files (Claude and OpenAI do; most local models do not: export a CSV or upload a screenshot instead). Dates must include the year.

Agents can do the same over MCP: match_statement with the payments, then set_charge_status for each match after showing the user.

Read the dates of a contract

Attach the contract, the terms or an order confirmation to a subscription (Documents on its page), then press Read dates. The model reads it and suggests:

• the notice period (in days; a month counts as 30 days),
• the page where it is cancelled or managed,
• when it started, and when a free trial ends,
• the account or customer number it is under.

Each suggestion shows the words of the document it comes from, and what Exhale has now. Untick what is wrong. Exhale checks the rest as an edit, shows the last day to cancel it would give, and saves only when you press Apply. The model has one tool, to report what it found, and cannot change anything; whatever the document says, the worst it can do is suggest a wrong value for you to untick.

Limits: PDFs up to 10 MB (the size of an attachment), pictures up to 5 MB. Without the assistant, the button is not there; documents can still be attached, downloaded and deleted.

Limits and costs

You pay your provider for what the assistant uses. To keep that bounded, one message calls the model at most 6 times, prepares at most 5 changes, and each member can send a burst of 12 messages and then one every 15 seconds.

To see what it costs:

• Activity → Assistant lists every question and every file a model read: who, where, the model, how many calls, the tokens and an estimated cost. Owners see everyone’s; others see their own. Never the words.
• Settings → Built-in assistant shows this month’s total.

The cost is an estimate from a price list (US dollars per million tokens) that Exhale knows for Claude models. For another model, or if your price differs, type it under Price in the same settings; for a local model type 0. A model without a price shows tokens only: Exhale never invents a cost. The server log has the same numbers (assistant usage … input_tokens=… output_tokens=… cost_usd=…).

When it does not work

It says	Meaning
The model provider rejected the API key	The key is wrong, revoked or out of credit. Fix it in Settings
The model set in Settings does not exist…	A typo in the model name, or the key cannot use that model
The model did not answer	Timeout or the address cannot be reached. For Ollama: is it running, is the address right from where Exhale runs?
The model provider is asking to slow down	The provider’s rate limit. Wait a minute
That is a lot of messages in a short time	Exhale’s own limit per member
The model declined to answer that	The provider’s safety filter refused the request

Check after setting it up

The assistant is tested with a scripted model; it has not been run against every real one. With a new provider, check these once:

1. Test in Settings says the model “can use Exhale’s tools”.
2. Ask “what do I pay per month?”: the number matches Overview.
3. Ask to add a subscription: a card appears, and nothing exists until you confirm. Confirm, then Undo.
4. Ask to cancel one: the card is marked as destructive and names the right subscription. Discard it.
5. Import a screenshot: the list matches the picture.
6. Ask “show me all my subscriptions”: a screen lists them, and the answer appears as it is written.
7. Activity → Assistant shows the question with its tokens, and a cost if the model has a price.

Developers can run the eval set against a real model: EXHALE_EVAL_API_KEY=… go test ./internal/assistant -run TestEval -v.