How Brassdeck works

What it is

Brassdeck holds everything you have started on one canvas, one card per idea, and keeps an honest record of where each one actually stands. It is not a list of things you owe: a card is an intention worth keeping, and stopping one on purpose is an outcome the app is built to make easy.

There is no ranking anywhere in it, no red, and no count of what you have failed to do.

The words

Eight words, each meaning one thing. The app uses them the same way everywhere, and so does the assistant when you connect one, because the connector is handed this list before it does anything.

Two of them are worth pinning down. The canvas is the surface and the board is the content: you pan the canvas, and you share the board. And a stage says how far an idea has travelled toward being real, which is a different question from whether you touched it today.

The six stages

Four of them are a journey. The last two are exits, and they are not failures: they are kept, so the decision survives the idea.

Parked and dropped are left out of every count in the app. Counting them would quietly turn your own decisions back into obligations.

The honest clock

Every card carries one number: how long since it last actually moved. Most tools lift a card to the top the moment you rename it, reorder it or drag it somewhere neater, so the list flatters you and the number stops meaning anything. This one is kept separately and moves only when something real happened.

A card dragged across the canvas after 147 days still reads 147 days. That is the whole point of it, and this is the full list rather than a rule of thumb.

Parking or dropping counting as progress is not a trick. Treating a deliberate stop as neglect would be the same as shaming it.

Fading, not nagging

A card nobody has moved in a while recedes. Its paper tints toward the desk, the way paper left out does, rather than turning a colour meant to alarm you. The writing on it stays as readable as on any other card. The information is identical and the feeling is not, which is the entire argument.

• The fade is scaled to each card's own cadence, not to a fixed number of days. A quarterly idea untouched for forty days is fine. A weekly one untouched for forty days is not, and only the fade knows the difference.
• A cadence of zero means never. Set it and that card never fades, however long it sits.
• Parked, dropped and finished cards never go cold at all. They already stopped, on purpose. Counting the days since you last touched something because you finished it would be the meanest arithmetic this app could do.

Waiting on someone

Open a card and under the four stage buttons is a line that asks who or what you are waiting on. Write a name, a company, a queue. The card then says Waiting on Wise support · 11d on its face, and the day count is the half that matters: a label tells you what you already knew when you set it, and a number is a reason to go and chase somebody.

Waiting is not a stage, on purpose. A card waiting on a bank is still at building, and the day the bank answers it should still be at building rather than having to guess where it was. So the stage still says how far the idea got and the wait says who is holding it, and you never lose one to record the other.

A waiting card is no longer called stalled, because stalled means nobody has decided what happens next and here somebody has: they reply. The clock keeps running, though, and that is deliberate. A wait nobody has chased in six weeks is the single most useful thing this app can point at.

Neither starting nor ending a wait counts as progress. Marking it is not work, and the reply landing was somebody else doing it.

The star

A star means "I am on this right now". It is one bit. There is no second level, no ranking behind it, and starring a card is not progress: it does not touch the clock.

Star anything and a Starred row appears in the views menu, showing only those cards. It is built in rather than a view you made, so there is nothing to name and nothing to keep in step. Parking or dropping a card clears its star, because being on something you have deliberately stopped is a contradiction.

The room

Open a card and the panel offers Mark it done. Finished is not a seventh stage, deliberately: how far an idea travelled and whether you are done with it are two different questions, and they genuinely disagree. A thing can be live and still being worked on. A thing can be finished having never left building.

• Where it goes. A finished card leaves the working canvas for the room, a fixed place further down, laid out on a grid in the order things were finished.
• How to get there. The views menu, beside Tidy. It is a place rather than a filter, so you travel to it.
• How to get back. Escape, or the browser's Back button. The room pushes a history entry so Back does the obvious thing.
• What it looks like. Faded brass, never grey. A finished card keeps its accent and loses saturation. Draining the colour would file it with the cold ones, and it is not cold. Its meta line shows the date it was finished, where the clock used to be.
• Reopening. Nothing is deleted. Reopening puts the card back where it was on the canvas.

Finishing a card is for members of the workspace. Someone holding an editing link can tick steps and write, but cannot make a card vanish from a link somebody else is reading.

Views

A view is a lens, not a folder. A card can be in several of them and it never moves to get there. Switching is instant because everything is already on the page.

• All cards is the default, and Starred appears once you star something. Views you make sit under those, each with its count.
• Tidy the canvas lives in the same menu. It rewrites the position of every card you own, and the toast that follows offers "Put them back".

Three named pages ask a different question of the same board: Focus for how the load divides right now and what is actually moving, Patterns for why things stop, counted across every stall, and Everything for the whole board as one list ranked by how long each thing has sat.

Sharing

A share link is a bearer credential: whoever holds it has the access it was made with, and there is no sign-in behind it. Everything below is chosen when you make one, and can be changed on a link already handed out.

• Viewer or editor. A viewer can open any card in the link and read all of it, steps and history included, and change nothing. An editor can tick steps, write them and change what a card says. Someone opening an editing link is asked for a name and a colour, so you can see who changed what. That name is whatever they typed and is never verified.
• Scope. A link covers the whole board, or one view, or a single card. A view-scoped link shows only the cards in that view, and it says so on the link itself rather than looking like the whole board.
• Expiry. Never, unless you set one; a week and a month are there for a link you want to stop on its own. A link that has expired or been turned off shows the person a page saying so, with a way to ask you for a fresh one.
• Two things are held back unless you turn them on. Both default to off: Stall reasons, the reason you gave when something stopped, and Dropped cards. They are the most personal things in this app and they do not travel by default.
• Revoking is permanent. A revoked link is never revived, and it stays on the page with its view count, because how often something was opened is part of the record. Only a hash of the link is stored, so nobody, including whoever runs the app, can read a live link back out.

Ask

Ask is the assistant inside the app, in the toolbar and on the phone's tab bar. It runs the same tools the connector does, against your board, and it streams its answer as it works so you can see which tools it called.

You can also talk to it from inside a card. A line that begins with @chit in the step box or in the writing is a message about that card: "@chit break this into five steps" writes the steps, "@chit which mic under $100" answers, and the answer is kept under your line in the writing as "Chit:".

This is the opposite direction from the connector. Ask runs a model on your behalf and is billed per message. Connecting Claude or ChatGPT is not.

Routines

A routine is a standing notification: when something on the board reaches a number you chose, the workspace gets an email. You set one up by asking for it in plain words rather than filling in a form, and every routine fires once and then stops. A notification that arrives every night is a feed, and a feed gets muted.

Every routine is listed under Connectors with the number it is watching for and a way to stop it, whoever or whatever created it. The email quotes your own request back at you and carries nothing else.

What a step knows

Write a step that is nothing but a link or a name, and the app asks the world about it. There is nothing to connect and no key to paste. The rule is the same for all three: the whole line has to be the thing, because "look at that repo later" is a step and putting a reading on it would be answering a question nobody asked.

• Domains. A line that is a domain name shows whether it is still available. A label after a dash or a colon is fine.
• YouTube. A video link shows its views, and whether they have moved since the last look. A nightly snapshot is what turns a number into a direction. This one needs a key held by whoever runs the app, so it can be switched off; the Connectors page says which.
• GitHub. A repository shows how long since anyone pushed, which is the only number here that gets worse on its own if you do nothing. And a push moves the card: each night, a repository that was pushed to since the night before counts as real progress on every card that links it, so a card whose work is a repository is never called stalled while the commits are landing. A pull request or an issue shows what became of it, and when the pull request merges or the issue closes, the step that linked it is ticked for you. Parked and dropped cards are left alone.
• A calendar. A step that is a calendar address (the private .ics address from Google Calendar, a TripIt feed, an Apple published calendar, any webcal link) shows the next thing on it and how far away it is: "Flight to Dubai · Fri 3 Oct · in 12 days". That is how a trip, a launch or a deadline gets its date on the card. Nothing on a calendar moves the clock; a date arriving is not you doing something. The address is read and never stored.
• Google. A Docs, Sheets or Slides link reads as what it is. Titles and last-edited need a Drive connection, which is not built.

Three card sizes

Every card is the same width and one of three heights: small, medium or tall. Drag the bar on a card's bottom edge and it moves between the three. There is nothing in between, and that is the feature rather than a limitation.

The alternative is a card whose size is set by whatever happens to be written inside it, so adding a sixth step makes it taller than the card beside it and no amount of tidying will ever line the two up again. Three sizes means a canvas you can actually arrange, and it puts the decision back with you instead of with the content.

A card holding more than it can show clips and fades at the bottom edge rather than stretching. Nothing is hidden from you dishonestly: the count in the bottom corner has always been the real number of steps and still is, so a card showing three of them still says 4/11.

All three heights are multiples of the canvas grid, so cards line up with each other and with the grid at the same time.

On a phone

A phone gets the deck rather than a shrunken canvas. An infinite canvas is a desktop idea: fit a real board onto a 375px screen and you land at twenty per cent, where a card is fifty-six pixels wide.

• Deck is the same cards and the same editor as a readable stack, ranked so the things asking something of you sit at the top.
• Board is one tap away and pinches to zoom properly.
• The raised round button in the middle makes a new card. It is the only control on the bar that makes something rather than showing something.
• More holds Focus, Review, Everything, Patterns, Share, Connectors and Settings.
• Install it. Add to home screen and it opens without browser chrome. On iPhone that is the share sheet, then Add to Home Screen.

Keyboard

On the canvas. All of these are ignored while you are typing in a field, except find and Escape, which belong to whatever is open.

While find is open

Cmd is Ctrl on Windows and Linux. Opening a card gives that panel Escape, so it closes before the room does.

How it connects

Brassdeck speaks the Model Context Protocol, and it is the server. Claude and ChatGPT connect to it. The instinct is the other way round, and which way it goes is the whole reason this is free:

• No API key is needed from you, and nothing is billed here.
• The model, the agent loop and the conversation all stay in the assistant you already pay for.
• One endpoint serves Claude, ChatGPT and anything else that speaks the protocol, rather than two integrations that drift apart.

It is JSON-RPC 2.0 over a single POST to a URL that carries your connection's token in the path. Protocol version 2024-11-05. The server advertises tools and nothing else: no prompts, no resources, no sampling.

Connecting

Sign in, open Connectors, choose Claude or ChatGPT, and decide before you press Connect whether it may edit. You are shown a URL once. Paste it into the assistant.

• In Claude: Settings, Connectors, Add custom connector, paste the URL.
• In ChatGPT: Settings, Connectors, Add, paste the URL.

One token per connection. Connect both and you get two, so disconnecting one does not sign the other out. Each row on the Connectors page shows a hint from the token, whether it can edit, how many calls it has made and when it was last used.

Read-only is not a refusal, it is a shorter list. A read-only connection is never even shown the tools that write, so the assistant does not try one and then report the refusal as though the app were broken. Of the 15 tools, 5 read and 10 write.

Every tool

Read out of the server itself, so this table cannot fall behind it. Names are what an assistant would ask for rather than what the database calls it. Inputs marked required are the ones a call fails without.

A tool that refuses answers normally and says why, rather than ending the exchange with an error. "Nothing here by that name" is a fact the assistant should read and act on, not a fault.

What it is told

The server does not only hand over a list of tools. Every connection is given a short briefing on the first call, so an assistant that has never seen this app does not invent its own words for it. In summary, it is told:

• The vocabulary, and to use it exactly. Card, board, canvas, step, stage, and the two exits. It is told which words it may never say.
• That this is not a list of things owed. A card is an intention worth keeping, and stalled means only that nobody has decided a next step yet, which is a question rather than a failure.
• That the clock is honest and must stay that way. Which actions move it and which do not, and that the number is reported plainly: never as a reproach, never with a target, never as a streak.
• That parked and dropped are kept forever. Never to suggest deleting them, never to count them as outstanding, never to call stopping something on purpose falling behind.
• That it is your board. If you name a card it has not seen, search before doubting you, and call a tool rather than answering from memory.
• Whether it may write. A writing connection is also told it may offer once, in one sentence, to put something on your board when you are clearly starting something with a life beyond the conversation, and to create it only if you say yes. Never unasked. A read-only connection is simply told it is read-only.

Tokens and revoking

• The token rides in the path. That is the one field every assistant has: you paste a URL. It is 32 bytes from a cryptographic random source, held to exactly the rules a share link is held to.
• Only its hash is stored. Nobody, including whoever runs this app, can print a live connection URL back to you. That is why it is shown once.
• Revoke is a column, not a delete. Disconnecting marks the row and keeps it, with the call count, because what once had access to your board is part of the record.
• Every tool is scoped from the token. The workspace and the member come off the token's own row. Nothing an assistant sends can name a workspace, which is the difference between an assistant that can read your board and one that can read everybody's.
• A connection is a key. Anyone holding the URL has the access it was made with. Treat it the way you treat a share link, and disconnect anything you are not using.