The source stays attached
#2 [11:40 · Lecture 7 slides · Safari] A model that does well on training data and badly on new data has overfitted.
Never explain your
project twice.思簿
Copy anything worth keeping and tap ⌥ twice — Ctrl on Windows. It lands in the project you are on, along with the thought that made you save it. Weeks later, hand the whole thing to any AI in one paste, and you can read every word it was told.
- Free
- Apple Silicon Mac · Windows x64
- Notarised by Apple on macOS
- Works offline
The background you keep typing again
Monday, you open a new chat and type it all out: what the project is, who it is for, what you already decided, what you tried and dropped. Wednesday, another chat — you type it again, shorter this time, leaving out the part that mattered. With Spool it was all saved while you worked. One click, one paste, and the AI starts from where you actually are.
Building something over months
The bug you already ruled out, the library you chose and the reason you chose it, the approach you dropped in week two. A fresh chat stops walking you back into circles you have already left.
A course, a thesis, a field you are learning
The line in the lecture you did not follow, the paper you will have to cite, the explanation that finally made it click. Come revision time, the AI quizzes you on your own material instead of the internet's.
Three AI chats a day, all starting from nothing
Different apps, different windows, and every one of them meets you as a stranger. Spool is the memory they do not have — and the same paste works in all of them.
Try it — no download needed
Do the whole thing right here
Save one note, pack it, and watch a brand-new AI chat pick up the exact point without being told the background. Then the last step, the one nobody believes until they do it: an AI asks to file something, and you decide. The content is written in advance; the flow mirrors the real app.
Or take a real one away. This is the actual output of one project's ⌘⇧P — instruction header, source labels and all. Paste it into whichever AI you use and ask it something about the course.
Copied — now paste it into your AI
Read the file first
About 8,000 characters — five saved notes and the thought the reader had about each.
How it works
Save it, keep it, hand it over
01 · SAVE
One key. Nothing to decide.
Copy anything, then tap ⌥ twice — Ctrl on Windows. It goes straight into the project you are working on, along with where it came from — down to the browser tab. The confirmation appears with the cursor already in a note box: type the thought that made you save this, or click away to skip. Your own words outlast the excerpt, and they are what an AI reads first.
02 · KEEP
A list, not a chat.
Everything you save stacks up in time order under one project. Two levels only — workspace, then project. No folders to build, no tags to pick. Come back next week and the newest notes show you exactly where you stopped.
A project holds more than lines of text: files you drop in stay with it, a date written inside a note gets flagged at the top of the project before it passes, and when something stops being true you can retire it or write the correction underneath. The old note stays where it is, with a line saying what happened to it. Nothing you wrote is overwritten, by you or by anything else.
03 · HAND IT OVER
One click, ready to paste.
Spool writes the project out as plain text: your own words first, things you read kept with their sources, anything an AI wrote marked so you can check it. Paste it into any AI — or just read it yourself to get back into the work. It is assembled on your own machine and put on your clipboard; nothing is sent anywhere, and the same project always comes out the same way.
This is what gets handed over.
You can read it. So can the AI.
Not a summary an AI made of your notes — the notes themselves, in order, each one still carrying where it came from. Below is a real project written out by Spool, unedited except where marked.
Your own words stay distinct
#4 [16:30] Revision plan: redo problem set 3 with the fee idea in hand, then watch lecture 8.
AI writing is signed and linked back
#5 [16:42 · Claude · MCP] Before Friday: problem set 3 question 2 is the overfitting one.
↩ cites #4
Open the complete real Pack example
# Project Context: Machine learning course Generated by Spool on 2026-08-12. 5 blocks total. ## How to Read This Context The blocks below come from FOUR different authority categories. Treat each category according to the rules in this section. ### 📖 Reference (authoritative) **Handling**: Treat as ground truth. Do not contradict. ### 🧩 Synthesis (already-formed understanding) **Handling**: These are someone else's synthesis. […] Do not treat them as facts. If they contradict Reference, defer to Reference. ### 🔄 Process (conversation traces — read for evolution, not facts) **Handling**: The literal content of these blocks is NOT a reliable source of facts. What IS reliable is the user's evolving questions. ### 💭 Personal (the user's own hypotheses and notes) **Handling**: Read these to understand where the user currently stands. If they contain factual errors, point them out directly — do not protect the user's feelings at the cost of correctness. [… the reading rules and the notation key run for 4,192 more characters …] ## Full Record (chronological) #1 [2026-07-29 10:20 · from course.edu · Safari] Week 6 is about overfitting. The homework wants a validation curve, not just an accuracy number. #2 [2026-08-02 11:40 · from Lecture 7 slides · Safari] A model that does well on the data it was trained on and badly on new data has overfitted. A bigger model is not the fix. note: This is the part I never followed in class. #3 [2026-08-02 14:05 · from AI chat · Safari] Regularisation is a fee charged for complexity: the model can still bend to the data, but every extra bend costs it something, so it keeps only the ones that pay for themselves. #4 [2026-08-03 16:30] Revision plan: redo problem set 3 with the fee idea in hand, then watch lecture 8. #5 [2026-08-03 16:42 · from Claude · MCP] Before Friday: problem set 3 question 2 is the overfitting one — that is the question the quiz will rhyme with. Do it with the fee idea, not with a bigger model. ↩ cites: [2026-08-03 16:30] Revision plan: redo problem set 3 with t…
The screenshots and the text above are made with a library built just for this, so nothing personal appears in them.
That is the whole loop. Install it, choose a project, and start saving.
Download for macOS Download for WindowsFree · Apple Silicon Mac (arm64), signed & notarised by Apple · Windows x64, not signed yet
And then it compounds
Week one saves you typing.
Week six saves you the thing you forgot.
Day one it already works: three notes, one paste, and you skip retyping the background. Six weeks later the same project holds the decision you no longer remember making, the link you would never find again, and the reason you ruled something out — and all of it goes across in that same single click. Nothing about how you use it changed. The pile just got deeper.
The app keeps the count for you. In the corner of the sidebar, what you have saved winds onto a spool: a hundred notes fill one, and it starts the next.
And it looks after you too
It knows where productivity and wellbeing balance.
Work long enough without a break and Spool locks its own window for five minutes: a countdown in the middle, everything else dimmed. One click ends it early, and if you do nothing it lifts by itself. How long a stretch runs is yours to set — thirty, sixty or a hundred and twenty minutes — and you can switch the whole thing off.
It only counts the time you were really there: Spool in front, and you having touched it. A few minutes in another app pauses the count rather than resetting it, and a laptop that slept all afternoon cannot hand you an hour you did not work.
Sixty minutes is the default, and the number is not ours. A 2026 study in the British Journal of Sports Medicine had 19,342 adults take a five-minute movement break every 30, 60 or 120 minutes. Every 30 minutes did the most for fatigue and mood, but too few people could keep it up. Every 60 minutes was the one they stayed with — and their work did not suffer for it. Diaz et al., BJSM 2026
MCP — beyond pasting
Paste it across,
or let the AI reach in.
Both routes end in the same place: an AI that knows the project. One works immediately in any browser tab. The other you set up once, and then stop copying altogether.
Route A · nothing to install
Pack it and paste it
One click, one paste, any AI, any tab. This is the whole product — nothing is held back from people who stop here, and it works on a machine with no AI app on it at all.
The fetching is yours to do.
Route B · connect it once
Let it read your projects
With your permission, the AI app you already use opens Spool itself: reads weeks of context, searches every project, and files what it finds back in the right place. You stop copying and start asking.
One switch in Spool's settings, and a restart of that app.
Apps you chat with
Claude DesktopCodex
You ask in the chat box, the same as always. In the ChatGPT app that means a Codex conversation — an ordinary ChatGPT chat runs on OpenAI's servers and cannot reach your machine.
Tools you code in
Claude CodeCursorVisual Studio CodeWindsurf
You ask in the editor's AI panel, or in the terminal.
Either way you stay where you are — Spool just hands the notes over. Spool can write the MCP setting for those six from one button. Restart the app so it can load that setting; Spool then shows whether it has actually connected. Not on the list? Copy one short setting out of Spool and paste it into your app. How to actually use it lives in one place, next to that button: a short paragraph you can paste to your AI — or just read yourself.
When an AI writes back
It signs its name, and the risky moves wait for you.
An AI you connected can add to a project — that is most of the point of connecting it. What it writes carries the name of the app that wrote it and sits below your own notes, never over them. It cannot edit one of your notes, and it cannot delete one.
What stops and waits for you
-
The moves that touch old notes wait for you
Saying that an older note has been overtaken, splitting one long passage across several projects, and asking to read a file you dropped in — those three stop in a queue instead. You read them, and you keep them or throw them away.
-
It corrects, it does not erase
When something you saved has been overtaken, an AI can point at the one sentence it corrects. The old note stays exactly where it is, with a line underneath saying what corrected it.
- Files have their own switch. A document you dropped into a project is not readable by an AI until you say it is, file by file.
- Retiring something stays yours alone. You can mark a note as no longer valid; it leaves the packs but stays in the library and stays searchable, and the pack says out loud that something was left out.
- The switch is one click away, both directions. Reading is off until you turn it on. Writing is a second switch underneath it, on by default once reading is — turn it off and an AI can read your projects and change nothing.
While you are not looking
Tell it what to watch. It stays quiet until something changes.
Write a few plain lines about what would actually matter to this project — a decision you are waiting on, a page that might change, a number worth rechecking. Spool hands those exact lines to a command-line AI you have already installed, it goes and looks, and whatever it finds waits in the same queue as everything else. When there is no news, it says nothing at all. Separately, Weekly Review can look across every project at once; it has no web-search tools and lives in its own review screen.
This is the only Spool action that receives web-search tools — and only against the lines you wrote. Weekly Review and goal drafting get no web tools, though they still pass the context they need to your CLI provider. Capture, organise, search, and Pack stay local and work offline. The engine is off when you install Spool, and it turns off where it turned on.
Before you download
What you are actually installing
The permissions macOS asks for, and what they are for
This is the step that surprises people, so here it is up front. Spool uses Input Monitoring to hear the system-wide double-tap of ⌥. Accessibility is optional, and does two things. It removes the second key press after Spool captures, so another app using the same gesture does not open at the same time. It is also what lets the confirmation take the keyboard, so you can type your note without clicking into it first.
The first time you capture something out of a browser, macOS asks once more, for permission to talk to that browser. That one is what lets a note keep the tab you were reading instead of just the name of the app.
Decline Input Monitoring and double-tap ⌥ is unavailable, but a custom capture shortcut still works. Decline Accessibility and capture still works, though another app using the same gesture may open too, and you have to click into the note box before you can type. Decline browser Automation and capture still works, with the source falling back to the browser name. None of these choices makes Spool send anything anywhere on its own.
On Windows there is nothing to grant. The double-tap of Ctrl is read through Raw Input, which asks for no permission and is not a keyboard hook — the thing antivirus software is right to be suspicious of. Source capture there is the name of the app you copied from; the browser tab title is a macOS-only detail.
Nothing is tracked, and nothing leaves unless you switch it on
No analytics, no telemetry, no crash reporting, and Spool's own process cannot make a network request — it is built without an HTTP client. Content leaves your machine only along a route you switched on: paste a Pack into a service, enable an MCP client, ask a command-line AI you installed to draft follow-up goals, run Follow Up, or write a Weekly Review, or turn on the API engine so Spool can compress a project for you. On the first three the receiving program controls what happens next; on the API engine the request comes from a separate bundled subprocess, with a key you supplied and on your own quota. Your clipboard is read at the moment you use the capture shortcut and not otherwise.
One file, on your own machine
Every workspace, project, note and annotation lives in a single spool.db
file. Copy it, back it up, delete it. No account to create, no server to trust, nothing to
cancel.
Notarised on macOS, not yet signed on Windows
The Mac download is signed with an Apple Developer ID and notarised by Apple, so macOS opens it without the "unidentified developer" warning. Apple Silicon only — an Intel Mac will download it and fail to open it. The Windows installer is not signed yet, so SmartScreen stops it the first time: "More info", then "Run anyway". That is worth knowing before you download rather than after. New versions are fetched by hand on both, which also means the app never goes online to check for one.
Made by one person — including the day I wiped my own database
No company, no funding, no roadmap meeting. The build is written up as it goes, and that write-up includes what happened that day and what I changed so it cannot happen again. The build log is public; the source is not open at this point.
Read the full privacy policy → How it was built — and what broke →
Questions
The things people ask first
Will it run on an Intel Mac?
Not today. The build is Apple Silicon (arm64) only — on an Intel Mac the download succeeds and the app will not open. There is no Intel build to point you at.
Do I have to connect an AI at all?
No, and nothing is held back if you never do. Packing a project and pasting it works with any AI in any browser tab, on a machine with nothing else installed.
Windows? iPhone?
Windows, yes — 0.5.0 is the first release for it. It is x64, capture is a double-tap of Ctrl instead of ⌥, and the installer is not signed yet, so SmartScreen warns you once. iPhone, no — there is no version, and I would rather say that plainly than put up a waiting list for something that does not exist.
Can an AI change or delete my notes?
No. It can add a clearly AI-signed note to a project. If it proposes that one sentence in an older note has been overtaken, that proposal stops in a queue you approve. Marking something as no longer valid, and deleting, stay yours alone. Nothing an AI writes can appear as if you wrote it.
Do I need an API key or a paid AI plan?
Not for the everyday use of it. Reading through an AI app uses whatever plan you already have in that app, and the watching runs a command-line AI you have already installed and logged into, on that subscription's quota. One optional feature does take a key: the API engine, which lets Spool compress a project, or check it for notes a later note has overtaken. It is off until you switch it on and paste a key, and each run spends your own credit.
Why is it free?
There is no server, no account and no company behind it, so there is almost nothing to charge for. One person builds it, and it runs entirely on your own machine.
What happens to my notes if you stop working on it?
They stay exactly where they already are: one spool.db file on your own
machine, in SQLite — a format plenty of other software can open. Spool going away does not take your
notes with it.
Is it useful if I do not use AI much?
Yes, though you would be using half of it. Packing a project into readable plain text is worth it on its own — reading it back is how most people restart work after a week away.
How do I update it?
By hand for now: download the new dmg, or the new Windows installer, and replace what you have. There is no automatic update channel yet, which also means the app never phones home to check for one.
Does it watch my clipboard all the time?
No. Spool reads the clipboard at the instant you double-tap the capture key — ⌥ on macOS, Ctrl on Windows — and at no other time. It does not poll, and it keeps no history of what you copied.
Do I have to organise anything?
There are two levels — workspace, then project — and both already exist when you open the app. No folders, no tags. Captures go to whichever project you marked as the one you are working on.
The mark: a spool seen from above, with its thread pulling free.
Stop explaining the same project again.
Free, works offline, and needs no account.
Download for macOS Download for WindowsApple Silicon Mac (arm64), signed & notarised · Windows x64, unsigned · no account · nothing is tracked · all releases on GitHub
After you have used it
Tell me how it went.
Spool has no account and no telemetry — nothing about how you use it ever reaches me, which also means it cannot tell me anything either. Whether it worked or it broke, I only find out if you say so.
Two minutes. Only the first one is required and you can skip the rest. This page has no server and collects nothing itself, but the survey is not this page: it is hosted by SurveyMars, and what you write there is stored on their servers. To keep it between the two of us, use the email below instead.
Or write to me instead
Or skip both: email me directly · open an issue on GitHub · watch the releases page to hear about new versions.