Five things sit above the menus, because they’re where I start.

Home — what’s actually open

Not a dashboard of vanity numbers. The front page is a worklist: up to three things I’ve said lead the day, counts of what’s waiting (and each one is a link into the list behind it), what landed recently, and anything I’ve pinned there myself — dragged into the order I want.

Two small honesties worth noting. The greeting matches the actual time of day, and "due today" uses my own calendar day rather than a clock instant, so a task due today isn’t counted early in the morning and doesn’t drop off before the day is over.

Chat — talk to the brain

A conversation with everything Station knows. Three settings decide how much it may reach for: Auto (it judges per message — the default), On, or Off for a plain conversation with no reach into my data at all. Whatever I pick is saved onto that conversation.

Every reply carries badges that are information, not buttons: ● Claude or ◐ local model for who answered, ⛁ grounded if it used my own material, and whether an attached image was actually looked at. Closing the tab mid-answer doesn’t cancel it — the reply keeps generating and reattaches when I come back.

Chat cannot run commands on my machines. Anything that has to touch a host goes through Health, which asks before it changes anything.

Ideas — the whole corpus

Every captured idea in one place, filtered by mode and by state.

Inbox — the mail gets read before I do

Station reads what arrives, suggests what to do, and acts on its own for exactly two verdicts: delete and decline — and even then only after judging that the message is the routine version of something it already understands. Anything unusual is downgraded to a suggestion and left alone.

Each row carries a confidence — ● high · ◐ medium · ○ low — and the word beside it says where the suggestion came from: a rule I wrote, the local model, or a school-portal match.

The one that catches people is Acknowledge versus Dismiss. Both clear the item. Acknowledge teaches Station nothing; Dismiss counts against ever automating that kind of message. Use Dismiss when you want less of something.

Nothing happens invisibly, either: everything Station handled on its own is listed on Done for you, which is the screen below.

Search — one box over everything

Keyword and meaning together, across conversations, the engineering journal, ideas, plans, knowledge, memory, tasks, the homelab inventory and the wiki. Results are grouped by kind, with the matched words highlighted.

Two things I like about it. regex turns the box into pattern matching. And when a search finds matches inside a work project, it says "N more matches are inside workspaces" and gives me the workspace name and a count only — never a title, never a snippet, because showing those would defeat the boundary the line exists to explain.

(No screenshot of this one. Search puts every topic I’ve ever used into a filter list, and one of those is a client’s name — so the honest thing was to leave the picture out rather than publish a page I can’t fully vouch for.)

The quick-create menu: Capture for an idea, plus direct doors to a new note, a new task, and a recording, and Triage for deciding what the captured things become.

Capture — one box, three levels of pushback

This page always creates an idea. The three buttons underneath decide how hard Station pushes back, not what gets saved: Park, Triage (one question), or Dig (up to three). It saves immediately either way and thinks afterwards.

Triage — decide what each thing becomes

The decision queue, sorted by what’s been waiting rather than by what’s loudest. Keep and challenge, promote to a plan, send to research, or bin it.

Capture has enough depth for its own page — the other front doors, the shortcuts, what "completeness" actually measures, and the ways things arrive without me typing at all. See Capturing things →.

The largest menu, and the one I live in.

Work — today, this week, active plans

A read-only summary: everything due by the end of today and unfinished, everything due in the next seven days, and every plan still marked active. The counts are totals even though the lists are kept short.

One thing worth knowing: like the calendar, this page is built entirely from due dates, so a task with no due date appears in neither column, however important it is.

Plans — the arcs I’m on

Every plan with a live progress bar. A plan’s priority and due date are borrowed from its unfinished tasks, so both move on their own as work completes.

Inside a plan, Board and List are two views of the same tasks, and a Gantt strip underneath is drawn from whatever has a due date.

Tasks — a board you can shape

Standalone tasks get a drag-and-drop board. Group by switches what a column means: Bucket (columns I made up myself, plus a "No bucket" column that always exists) or Status (not started, in progress, completed). Deleting a bucket column doesn’t delete its cards — they move to No bucket.

Underneath, In plans is a flat list of every task that lives inside a plan, so nothing hides in a plan I’d forgotten about.

A due date is a day, not a time. There’s no time picker anywhere, and a task due today stays out of "overdue" for the whole day.

Calendar — due dates, by month

Despite the name this is not a diary: there are no meetings or appointments on it, only tasks with due dates. Clicking a task opens its editor over the calendar, so the page reads like a calendar and edits like the task list. On a phone the month becomes a date-by-date agenda rather than seven squeezed columns.

Meetings — the commitments I make out loud

Meeting audio goes through Whisper locally and Station reads the transcript looking for one thing: first-person commitments I made. Not everyone’s action items — mine. Each candidate waits for a confirm or reject, and every decision trains the extractor.

Since 28 August 2026 a recording knows which meeting it came from: Station mirrors my work calendar, reads the attendee domains, and files the recording against the right project — badged internal, external or unknown, with the domains that decided it. Two one-click corrections are available, and the matcher never overwrites them.

(The swear jar lives here. It’s a persistent per-word tally that survives the recording being deleted, which feels like a design decision aimed directly at me.)

Debriefs — the evening call

Every weekday evening Station calls my phone and interviews me about the day, in a clone of my own voice. The review page is where those sessions live.

Workspaces — one container per effort

A workspace holds everything about a single subject: chats, notes, files, meetings, plans, tasks and a journal. The type matters — a general workspace gets items, calendar, journal and chat; a work project or client workspace also gets plans, tasks, files, notes, meetings and suggestions.

The important idea is the boundary. Anything created inside a workspace is ○ workspace-only and stays out of my global lists until I explicitly share it with Station, at which point it reads ● shared with Station and keeps its place in the workspace too. A workspace can be a scratchpad without cluttering everything else.

Client work has its own page — the register of what’s actually settled, the decks, the meeting matching: Work →.

Wellbeing

Weight, medication, meals, water, getting outside, and the rhythms of my day mined from the house’s own sensors. It’s the most personal screen in Station, so it gets its own page rather than a screenshot here: Wellbeing →.

Draft Studio — write in my voice

Give it a topic and a brain-dump; Station grounds a brief in my own posts and talk transcripts, and a coding agent drafts the piece into a private repository, one branch per draft, each revision a commit. Then it reconciles the draft against what I actually published, and the edits I made by hand feed back in as voice signal.

Development, Engineering Journal and GitHub

Three screens about Station building itself, all new since July.

Development is the board that tracks Station’s own work, with a session review that reads back the week’s coding sessions and tells me what keeps going wrong. Engineering Journal is a readable diary of those sessions, grouped by day and project. GitHub watches my repositories, the dependencies that actually break my code, and — since 14 August 2026 — tells me when somebody has sent me a pull request.

How Station ships itself is a page of its own: Station builds itself →.

Files — and a reader, not a download

What used to be "Documents". Since 24 August 2026 a file opens in a reader before it downloads — PDFs, images, text, Markdown, Word, PowerPoint and Excel — with a visual mode and an extracted-text mode. Rendering stays offline and sandboxed. Download is still there for every format.

Bookmarks, Notes, Sources and Research

Bookmarks are saved links with a note on why I kept them, an AI synopsis, and a weekly digest of the ones whose pages have changed underneath me. Notes are raw observations with deliberately no model passes — sometimes I want to write something down without being interrogated about it. Sources are the canonical things I cite, each saying what it teaches: voice, facts, or neither. Research runs long deep dives locally and hands back a report I can pin to an idea or a plan.

Memory — what it believes about me, and how sure it is

Short, structured facts about me, my projects and my people. Every fact carries a confidence and a validity window: it can be forgotten or superseded without being deleted, so there’s a real history of what Station once believed rather than only what it believes now.

Nothing enters memory silently. New facts land as pending, and since 3 September 2026 they come and find me on the Follow-ups page as ordinary sentences with ● Accept and ○ Reject beside them.

Graph, Entities, Topics and Positions

Graph is the links Station found on its own between ideas, entities and the cross-domain surfaces — rebuilt every night, and filterable by kind and by how confident it is. The thing I half-remember is usually two hops from wherever I’m looking. Entities are the people, companies and things it has extracted. Topics are the recurring threads, with a volume chart of what I’ve been chewing on. Positions are the things a project has actually settled on.

(No screenshot. The graph is a list of the real entities Station has extracted from my life, and several of them are people and places rather than software.)

Wiki — Station’s own living docs

The architecture notes, decisions, roadmap and changelog, rendered and editable in-app. Edits land on disk, get picked up by version control, and feed back into the knowledge index so the model knows how Station itself works.

On 10 August 2026 the homelab stopped being a subsection of "Browse" and got a menu of its own — sixteen screens about the fleet Station watches over. Highlights:

Homelab map — every host, stack and container

Seeded straight from the real environment, with a generated synopsis for each thing. Because it’s all just more nodes in the graph, I can ask "which of my ideas relate to this service?" and get a real answer.

Homelab Chronicler — the why layer

The map knows what’s running. The Chronicler knows why it’s set up that way. It interviews me about a service, keeps the durable reasons separately from the live facts, and updates the facts on its own when things move — asking for a review only when something looks genuinely retired or has changed role.

(No screenshot here either — the Chronicler’s index page carries enough of my own prose about the house that one personal term survived the redaction pass. The homelab module page has the screens that are safe to show.)

Container updates — a board you can work

Not a list of version numbers. Each row says what an update actually means for how I use the thing, and stops nagging once the container is current.

Health and backups

Health is where a host problem gets investigated — a conversation that works over the network and asks before it changes anything — and it now carries an always-visible sign-ins panel, because health has to be asserted rather than inferred from an absent alarm. Backups collects every backup job into one schedule-aware history, audits encryption and retention, and escalates anything failed, overdue or unknown.

The whole fleet, and the agents that watch it, have their own page: Homelab →.

Agents — autonomous work that reports back

The scheduled agents that do, not just think: a container-image monitor, a homelab groundskeeper, a backup guardian, a diagnostician, a telemetry watcher. Each shows when it last ran, what it found, and what it actually did — with severity written as a shape and a word, never a colour on its own.

Hermes — the overnight thinker

The thing I called "someday" in July. Since 1 August 2026 Hermes reads the day’s thinking while I sleep and has something to say in the morning. It’s suggest-only — it can look, not touch — and a thought worth keeping can be promoted into a real idea.

Jobs, Downloads and Stats

Jobs is the scheduler in the open: every background job with its schedule, last result, duration and a "run now" button. Downloads tracks how the things I publish are being used. Stats counts what Station holds.

Each agent, what it may touch and what it reports: Helpers and agents →.

Station’s app-of-itself menu.

Follow-ups — features that have to earn their keep

Every shipped feature is registered with a test plan and brought back at 2, 7, 14 and 30 days: did you use it, and is it doing what you wanted? The check-ins are grounded in a local usage log — nobody’s analytics platform — and drip-fed one a day so it never becomes homework.

The page also carries Needs a look: a block that appears only when a background job or a knowledge source keeps failing, each row with ◐ warn in shape and words, a one-line reason, and a link to where it can be fixed. When there’s nothing wrong, the block is completely absent.

Progress — how Station has grown

The in-app roadmap and changelog. As at 5 September 2026 that’s 848 changelog entries and 277 architecture decisions since the first commit in April.

Settings, Security and the rest

Settings was seventeen open sections; since 19 August 2026 it’s seven collapsible groups. Security covers sign-in: a password, mandatory two-factor, recovery codes, and passkeys — every page in Station is closed unless you’re signed in. Recycle bin holds everything user-created for 30 days with a visible "expires in N days" countdown. Feedback is where I file gripes about Station from inside Station, and since 13 July 2026 a piece of feedback gets interrogated into an actual development item rather than sitting in a list.

The recycle bin is the one below — every user-created thing, held for 30 days with a visible countdown. (Settings isn’t shown: expanded, it’s a page of API keys and connection strings, and no amount of masking makes that a screenshot I’d publish.)

The newest thing on this tour is the least visible. Since 19 August 2026 every page has a written guide behind a ? in its corner, and a stale one can’t ship — the guide is part of the feature, not documentation written afterwards. Beside it is an ask box that answers questions only from those docs, so it can’t invent a feature that doesn’t exist.

Thirty-six of those guides exist as at 3 September 2026, and they were the main source for this tour. If a sentence on this page sounds suspiciously specific about what a button does, that’s why.