Skip to content

Try the demo

The repository ships a complete, invented household so you can explore every page of the app, and ask the coach, without a bank, a model or an API key.

Every screenshot and screencast of this documentation was made on this demo. Nothing in it comes from a real home.

What the demo household contains

The Rossi family: two adults, Anna and Luca, and two children, Mia (born 2012) and Noa (2015). The data covers the 24 months before today, so the demo always looks current.

Area What you find
Accounts 7 accounts: a joint account and a savings book at Banque Aurore, one account each for Anna and Luca plus Mia's junior account at Nova Bank, a rental flat account at Credit Horizon (all three banks are invented), and Noa's prepaid card brought in as a file import.
Consents One bank consent is close to expiry, so you see the warnings of the Connections page.
Transactions 24 months of salaries, groceries (Acme Grocers, Fresh Market), restaurants, energy and water bills, a summer holiday (AirVola, Hotel Miramare), pocket money and transfers between the family's accounts.
Loans and assets A mortgage (Homebank), a car lease (MobiLease), the family house, savings and a life-insurance balance.
Rental property A two-room flat let under a Pinel-type tax-incentive scheme, with its own loan (Lenderco), rent cap, tenant income limit and a vacancy between two tenants.
Subscriptions and contracts StreamBox, VideoMax, SoundWave, CloudBox, OldApp, TelcoCo, PixelTel, HomeSure, FitClub and more, several with a contract file, recorded cheaper alternatives and savings decisions.
Plans Budgets, goals (an emergency fund among them), open questions and memory proposals waiting for review.
Signals Alerts, coach insights and an AI usage history.
Logins A child login for Mia (she only sees her own money) and an adult login for Luca.

The Connections page of the demo household The Connections page of the demo household

The demo's Connections page: three invented banks, a consent about to expire and the prepaid card import.

Run it

You need a source checkout (Install, "From source"), uv, and Node with pnpm to build the web app once.

  1. Build the web app (once; the built app is not committed):

    cd web && pnpm install && pnpm build && cd ..

  2. Create the demo home in an empty folder:

    uv run python scripts/demo/seed_demo.py --home /tmp/coach-demo

    It prints the balance of each account as it goes. Everything (configuration, database, memory, secrets) is written inside that folder.

  3. Start the app on the demo home:

    scripts/demo/run_demo.sh /tmp/coach-demo

  4. Open the one-time login link the command prints (http://127.0.0.1:8799/login#t=...). You land on the dashboard as the owner, who sees everything.

  5. See what a child sees. In another terminal, print a link for Mia's login and open it in a private window:

    uv run python scripts/demo/login_link.py /tmp/coach-demo --user mia

The demo dashboard on a phone What Mia sees with her own login

Things to try

  • Ask the coach: "Why was July so expensive?", "Which subscriptions should we review?", "Are we OK next month?"
  • Memory > Proposals: the changes the coach suggested, waiting for a human.
  • Rental property: the scheme commitment and the tax-year candidates.
  • Kids' money and Who pays what: the household views.

The coach in the demo

The demo has no model. Its "coach" is a scripted stand-in for the claude command, put first on the PATH by run_demo.sh:

  1. It reads your question and picks the closest of a few prepared scenarios (a month that was high, the subscriptions, next month's forecast, or a three-month summary by default).
  2. It starts the real finance tool server and makes real tool calls on the demo data (explain_spike, subscription_audit, forecast, calendar, cashflow...). You see each look-up under "How this was answered".
  3. It fills a pre-written answer with the figures and evidence refs those tools returned.

So the redaction, the evidence refs and every number are genuine; only the wording is canned. No network, no API key, no cost. A question that matches no scenario gets the default summary.

The coach answering a question on the demo data The coach answering a question on the demo data

"Why was July so expensive?": the tool calls, then an answer whose figures come from the tools.

Safety rails

The demo scripts are written so they cannot touch a real home:

  • seed_demo.py refuses a folder that is not empty unless it carries the demo marker file it wrote itself (.coach-demo-home).
  • run_demo.sh and login_link.py refuse any folder without that marker.
  • Secrets use the file backend inside the demo folder, never your Keychain. Any COACH_* or Enable Banking variable of your shell is ignored while the demo is built.
  • The demo database is a plaintext SQLite file (insecure_plaintext_db = true). That is acceptable only because every row is invented.

Never point the demo at real data

The demo configuration turns off database encryption and uses a scripted coach. Use it for invented data only. For your own money, follow the Quickstart.

Rebuild it

The demo is anchored on today's date. To refresh it (or after pulling a new version), rebuild it in place:

uv run python scripts/demo/seed_demo.py --home /tmp/coach-demo --force

--force only works on a folder that already holds a demo. Another port: --port 8800 when you create it.

How the screenshots of these docs are made

scripts/demo/shoot.py drives a headless Chromium (Playwright) against the running demo app. It logs in with a one-time link of the demo home, then saves light and dark WebP screenshots of every page, a few phone-sized ones, Mia's own view, and with --video the short screencasts of the main flows:

scripts/demo/run_demo.sh /tmp/coach-demo &
uv run --with playwright --with pillow python scripts/demo/shoot.py /tmp/coach-demo --out docs/assets/screens

--only dashboard,coach limits it to some pages. Like the other demo scripts, it only works against a demo home.

The narrated film

scripts/demo/make_film.py turns the screencasts into the film of the home page, once per language (film.en.mp4, film.fr.mp4, with WebVTT subtitles and a poster). Title cards are rendered by headless Chromium. The voice-over comes from Kokoro, an open-weight neural speech model (Apache-2.0) that runs on your machine: no account, no key, and the narration never leaves the computer. Each scene lasts as long as its picture or its narration, and the voice is normalised to -16 LUFS.

brew install espeak-ng           # the phoneme library Kokoro needs (Debian / Ubuntu: apt install espeak-ng)
# download kokoro-v1.0.onnx and voices-v1.0.bin (model-files-v1.0 release of kokoro-onnx) into ~/kokoro
uv run --with playwright --with pillow --with kokoro-onnx --with phonemizer-fork --with soundfile \
    python scripts/demo/make_film.py --kokoro-dir ~/kokoro --lang en,fr

The narration and the title cards of each language are at the top of the script. --tts edge uses Microsoft Edge's online voices instead (through the open-source edge-tts client), when that service accepts the client.

See also