One desk for your mail, calendar and markets.

The assistant stays on your computer.

Watson is a desktop app for Windows and macOS. Your mail, calendars and market lists live in a database on your own machine, and Watson, the assistant, can read, summarise and draft on a model that runs right there, offline if you like.

Beta · Windows (64-bit) and macOS 11+ (Apple silicon and Intel)

Screens on this page are illustrations with sample data.

Everything that asks for your attention, in one place.

The CasebookWhat needs you today.

Your mail as a feed, day by day, with each email set as a post: who wrote, which account it came to, a line on what it's about, and the message itself. In Importance order the people who matter to you come first, and their posts are larger; mail that needs a reply is marked. Make it your start screen.

Runs on this machine

  1. Importance: who matters to you comes first
  2. Today: the feed, day by day
  3. Reply: how urgent the mail looks
  4. The email itself, in the post
  5. Priority: how much this sender matters

The Casebook, light theme. Illustration with sample data.

  1. The year as four season bands
  2. A day, connected to its card
  3. An event with a place gets a pin
  4. Click it: the map flies there and marks it

The Season view over a map, light theme. Illustration with sample data.

CalendarThe whole year on one screen.

The Season view draws the year as four bands around an open board. Works with Google, Microsoft, iCloud, Fastmail and any CalDAV server. The Join button appears for Teams, Zoom, Meet and nine more meeting services.

Put a map behind the board, and every event with a place grows a pin. Click it, and the map flies to the place and marks it.

Sends your own account traffic to your calendar servers, when syncing Sends the places you look up to a geocoding service, only once you turn the map on

PeopleA directory that writes itself.

Nothing is typed in. Watson builds it from the mail in every account you add, and sorts the people you write to from newsletters, receipts and robots.

When two addresses look like the same person, a work one and a personal one, it suggests grouping them and waits for you. Once grouped, that person has one history across all your accounts.

A person's page keeps their addresses, your notes, what Watson remembers about them and your whole correspondence. Their priority decides how high their mail sits in the Casebook.

Runs on this machine

  1. Sorted for you: people, robots, unsorted
  2. A merge it suggests, and you confirm
  3. Two addresses, one person, one history
  4. Priority: how high their mail sits in the Casebook

People, light theme. Illustration with sample data.

Switch off what you don't need. Nothing is deleted, only paused.

One bar, in plain words.

The search bar at the top of every view takes a sentence. Type “the contract Halden sent before the summer” and Watson reads it into filters you can see and remove: who, when, which words. When nothing matches exactly it says so, then shows the closest mail, or mail that matches by meaning rather than wording.

Runs on this machine, with the model you chose

  1. A sentence in the bar
  2. What Watson read into it, as filters you can remove
  3. No exact match: it says so, and matches by meaning
  4. Saved searches, kept as views
  5. In Stocks: a company by name, with its price

Mail search and a Stocks lookup, light theme. Illustrations with sample data.

Space then W, from any screen, opens a larger box for a longer question. The shortcut can be changed.

Save a search with the bookmark in the bar, and it becomes a view in the sidebar.

In the Casebook, Contacts and the calendar, the same bar narrows the view you are in.

In Stocks, it finds a company by name or symbol, with its price, and adds it to a portfolio or a watchlist.

An assistant that stays home.

Watson runs on a small model that downloads once and works on your own hardware, using the GPU when there is one: Wiggins, Lestrade or Mycroft, whichever suits the machine, or any GGUF model you choose. Prefer something bigger? Point it at Ollama or LM Studio on your network, add your own OpenAI, Anthropic, Gemini or OpenRouter key, or use your ChatGPT plan through Codex.

What it does

  • Answers questions about your mail, and shows the messages it read
  • Summarises long threads and redoes the summary when a reply arrives
  • Drafts replies (“Say yes”, “Ask for detail”, “Chase”) in a voice learned from your sent mail
  • Translates a message in place, with the original one click away
  • Tells you how your day looks: meetings, what needs a reply, who wrote
  • Answers about your stocks and the market from the figures the Stocks tab shows, and gives no advice
  • It never sends or forwards mail.
  • When an email's text has reached it, it won't save a memory without asking.
  • Each feature can use its own model. Keep incoming mail on the device while you chat with a cloud model.
  1. Lestrade, a small model on this device
  2. The messages it read, as rows you can open
  3. A reply, ready to insert

The assistant panel, light theme. Illustration with sample data.

Ask it the way you would ask a colleague.

Summarise a few messages

Ask for a summary of several messages at once. It searches, reads them, and says what is settled and what waits on you.

  1. Each step it took, in plain words
  2. The messages it read

Find the one email

Ask for a fact the way you remember it. It finds the booking and quotes the reference. A memory taken from mail waits for your OK.

  1. The answer, from the message it came from
  2. A memory from mail waits for your OK

Ask about a stock

Price, setup, news and the next earnings, read from the same figures the Stocks tab shows. It gives no advice.

  1. Read from the data service, step by step
  2. The setup's levels, in a table
  3. At work: short lines from the Holmes stories

It knows what is on screen

With a ticker open, “this one” means that ticker. The same holds for an open message, a contact, or the study you are writing.

  1. What you are looking at
  2. “This one”, understood
  3. An answer from the same score

Write a study

Describe a question in one sentence. Watson writes the study, runs it and fixes what fails; keeping it is your call.

  1. Written, run, fixed and checked
  2. Kept only when you confirm
Needs a larger model: a ChatGPT plan, a cloud key or a big local server

Mail and markets, one conversation

Ask about your stocks from Mail. The Mail side asks the Stocks side for you, and answers with both.

  1. The Stocks side, asked for you
  2. Mail and the market in one answer

The assistant panel, light theme. Illustrations with sample data.

It reads the small print so you don't have to.

  1. The red bar, in plain words
  2. A familiar name from an unfamiliar address
  3. A link that says one place and goes to another
  4. Remote images, blocked

A suspicious message, light theme. Illustration with sample data.

Security bar. A red bar explains, in plain words, when something is off: a familiar name from an unfamiliar address, a look-alike domain, a link that says one place and goes to another, an invoice.pdf that is really an archive. The checks run on your machine, and no AI is involved.

Runs on this machine

Remote images stay blocked until you allow them, so tracking pixels never report back.

Unsubscribe bar. One click for the link, and for a mailto unsubscribe you see the address before anything is sent.

Search. Full-text search across every account, in English, French, German, Russian, Armenian and more. It tells you when it has widened your search.

Runs on this machine

18:40 · sun 5° · light theme

Mail, a ticker and the news wall, light and dark. Illustrations with sample data.

Light for the morning. Dark for the late shift.

Watson comes with both themes. You've just watched one become the other.

Markets, set like a newspaper.

Market data comes from Watson's data service. Your licence key is the credential.

Sends your licence key and the tickers you look up to Watson's data service, while Stocks is on

News wall

The news wall turns your lists into an edition: a lead story, what moved, what to set aside.

  1. The edition's masthead and the day's reading
  2. Your portfolio's lead story
  3. Sources and the names they moved
  4. In brief

The news wall, dark theme. Illustration with sample data.

Ticker

Quotes refresh every minute, including pre-market and after hours.

  1. The after-hours price, and the close beside it
  2. Your list, on the same prints
  3. Candles, with target, breakout, support and stop
  4. The 50- and 150-day averages
  5. Volume and RSI, in their own panes

A ticker's dashboard, dark theme. Illustration with sample data.

Setup

A setup score from 0 to 100 shows its arithmetic. It is not a forecast and not advice.

Watches tell you when: up to six conditions over about seventy signals.

  1. The score, 0 to 100, and what was found
  2. Entry, stop, target, reward to risk
  3. What each part of the chart scored
  4. How the score was built, step by step
  5. Not a forecast, and not advice
  6. Tell me when: a watch and its conditions

The Setup tab and a watch, dark theme. Illustration with sample data.

Studies

Ask a question in plain words, and Watson writes the study for you.

A study is a small Python program that answers one question about a ticker. You don't have to read it: Watson writes it, runs it on real tickers, fixes what fails and shows you the answer. Keeping it is your call.

Studies run in a sandbox on your machine, with no network and no access to your mail. Writing one needs a larger model than the three that run on the machine: your ChatGPT plan, a cloud key, or a local server with a long context.

The study runs on this machine, with no network

  1. Your question, in one sentence
  2. Watson writes it, runs it, fixes it
  3. Answers on real tickers, filed with the version
  4. Kept only when you confirm
  5. Private to this machine, or shared

Write a study and the workbench, dark theme. Illustration with sample data.

Workings

Show workings runs a study again and lays out every step it took, beside its code.

Pick a step, and the lines that did it light up in the editor. Each step shows the prices it read and the values it worked out, under the study's own headings, so any number on a study's card can be traced back to where it came from.

  1. Pick a step, or scroll to it
  2. Its lines light up in the editor
  3. The prices it loaded
  4. The values it worked out on the way
  5. Run again, with every step recorded

Show workings in the study workbench, dark theme. Illustration with sample data.

Top scorers

Twice every trading day, at 07:45 and 17:00 New York time, the data service ranks about 1,300 US stocks by the same setup score.

  1. Two runs a day: 07:45 and 17:00 New York
  2. The score, and the rank it earned
  3. Sixty sessions, with target, entry and stop
  4. The analyzers, one line per leader
  5. The service's verdict, for and against

Top scorers, dark theme. Illustration with sample data.

  1. The video's tone
  2. The names, matched to tickers, with the day's move
  3. A summary and the points
  4. The transcript, with timestamps

Your channels, dark theme. Illustration with sample data.

What leaves your machine, and when.

Watson keeps your data at home, and it names every exception.

Every touchpoint that sends something off your machine, when it happens, and what is sent
TouchpointWhenWhat is sent
Your mail and calendar serversWhen syncingYour own account traffic
Watson data serviceWhile Stocks is onYour licence key and the tickers you look up
A cloud AI providerOnly for features you pin to a cloud modelThe text in that request
YouTubeWhile you follow channelsRequests for their feeds and video transcripts
Model downloadsOnce, when you pick a modelA request for the file, to Watson's server, or to Hugging Face for a model you fetch there
Sender picturesOn by default, can be switched offA hash of the sender's address, or the domain
Weather and mapsOnly if you opt inPlace names, to the Photon geocoding service; for the map, the area on screen, to OpenFreeMap
Update checksAbout once a monthYour app version and platform

Passwords, tokens and API keys stay in your system's keychain, never in the database and never in backups.

A model that fits the machine.

A hardware check reads your memory, cores and graphics and recommends one of three models, named from the stories: Wiggins for lighter machines, Lestrade for most, Mycroft for a machine with memory to spare or a graphics card. Downloads come from Watson's own server, resume after an interruption, and are checked against a SHA-256 fingerprint before first use.

The three are a recommendation, not a limit. Fetch any GGUF model from Hugging Face, or choose a file you already have, from an LM Studio or Ollama library or an external drive, and Watson uses it where it sits. Every model is tested on this machine before a profile can use it.

Sends a request for the file to Watson's server, once, when you pick a model
  1. Downloads from Watson's own server, checked
  2. Three sizes: Wiggins, Lestrade, Mycroft
  3. Chosen for this machine, and why
  4. Langdale Pike, for search by meaning
  5. Any GGUF: from Hugging Face, or a file you have

Settings › Models, dark theme. Illustration with sample data.

Recommended models: size, file size, memory needed to run, and notes
ModelSizeFile sizeMemory to runNote
WigginsFor lighter machines0.8B0.81 GB1.7 GBFor lighter machines
LestradeFor most machines2B2.01 GB3.1 GBFor most machines
MycroftFor a graphics card, or memory to spare3–4B2.57 GB5.5 GBFor a graphics card, or memory to spare
Langdale PikeFor search by meaning300M0.33 GBnot listedFor search by meaning
Any GGUF modelFrom Hugging Face, or a file you already haveyour choiceyour choiceyour choiceFrom Hugging Face, or a file you already have

Every model is tested against a sample mailbox before Watson gives it more to do.

Be first at the desk.

The beta opens in waves. Leave your email and we'll send a trial licence key when your wave is ready.

A licence key unlocks the modules you choose, and it is checked on your computer. Nothing about it is sent anywhere.

I use
Most interested in

One email to confirm, and one when your key is ready. This page uses no trackers.