Private Space
A local-first, self-hosted Notion alternative that runs in one Docker container — pages,
nested sub-pages, databases with table/board/calendar/gallery views, calendar feeds, full-text
search, voice capture — with a local language model that pilots your project boards and has
to prove what it did.
No accounts, no cloud, no telemetry. The whole workspace is one SQLite file and an uploads
folder on a volume you own.
docker compose up -d --build # → http://localhost:3000
Try it without installing anything
| |
|---|
| ▶ Pilot demo | Drive the board agent yourself. Give it a task, choose how the model behaves, and try to make it lie. No sign-in, no server, no network calls. |
| Overview & tour | Five capabilities, animated live in the browser. No video, no screenshots. |
| Build log | Eleven weeks of engineering history — hurdles, corrections and dead ends included. |
The demo's model is simulated (a static page cannot reach a GPU), but the evidence gate is
the real logic, ported from src/agent.js.
The idea behind the agent
A language model will move a card to Done without having done anything, and a board that lies
is worse than no board, because you act on it. So the guarantees are structural rather than
prompted:
- Done is reachable only through
finish_task, which refuses unless a tool that produces
something actually returned during this run. set_status cannot reach Done at all, or the
rule would be one synonym from bypassed.
- Evidence is the server's own record of what its tools returned — never the model's
argument. Asked for evidence after calling a tool, a small model types a plausible page id as
readily as the real one.
read_page deliberately does not count. Reading is an input, not an outcome.- There is no delete tool. Deleting someone's work is a human act.
- "Needs you" is a first-class outcome. That status existing is what makes the others
trustworthy.
- A completion is a claim until a person accepts it. The agent has no tool that writes the
review field, which is what makes it a gate rather than a label.
And the limit that prompting cannot fix, stated rather than glossed: the evidence gate proves an
artifact exists. It cannot prove the artifact was the right thing to make. Both models tested
wrote a document about a physically impossible task and marked it done. The review gate is the
mitigation; it is not a cure.
What is in the workspace
Editor — nested sub-pages, drag-to-reparent, block types (headings, lists, to-dos, toggles,
quotes, callouts, code, images, embeds), markdown shortcuts, slash menu, undo/redo, version
history.
Databases — typed properties including relation, rollup and formula; table, board, calendar,
gallery and timeline views; filters, sorts, saved presets, per-view column visibility, footer
aggregations, CSV import/export, sub-items.
Planning — a Today/daily planner, recurring tasks, reminders, a Pomodoro focus log, workload
analytics with overload warnings, and a subscribable ICS feed so deadlines land in Apple or
Google Calendar.
Capture — fully local Whisper transcription in the browser (audio never leaves the device),
voice quick-tasks from a phone or Apple Watch via Shortcuts, and a quick-capture inbox.
Collaboration — accounts and roles, page and row comments, a space huddle with threads,
@mention notifications, and read-only share links.
Automation — the board agent above, an autopilot across many project boards, work-log
ingestion from local git activity, and an MCP endpoint exposing the pilot to external tools.
Engineering
- 695 end-to-end tests across 117 Cypress spec files, 0 failing.
- CI runs on GitHub Actions as 12 parallel jobs, one per capability — a red job names a
broken feature rather than a shard number. The build refuses to start unless every spec
belongs to exactly one suite, because a spec that silently stops running is the one failure a
green build cannot show you.
- 269 commits, 53 merged pull requests, six numbered security audit rounds.
- Node 22 · Express · better-sqlite3 (WAL) · vanilla-JS front end, no framework · Cypress 13 ·
Ollama · MCP · web-push.
The build log covers the
hurdles in detail, including the ones that are usually left out — retracted measurements, a
container that was doing CPU inference for weeks next to an idle GPU, a fairness fix that was
unreachable dead code, and a field-name mismatch that blanked 54 of 72 pages while four specs
stayed green.
About this repository
This repository is the only public surface of the project, and it is open on purpose — read
it, fork it, or lift the gate logic out of demo.html.
| Open here | Kept private |
|---|
| The three pages on this site | The application source |
| The demo's agent gate, ported to run in a browser | The workspace database and uploads |
| The build log, including the failures | Deployment configuration and tokens |
Everything in this repository is MIT licensed — take the gate logic out of
demo.html and use it. The licence covers this repository only; the private application
source is not under it.
The application source lives in a private repository and is available on request.
Everything in this repo is static: three HTML files with inline CSS and JavaScript, no build
step and no framework. Web fonts are the only external request, and the demo makes no network
calls at all — no sign-in, no storage, nothing sent anywhere.
Contact
Happy to walk through any of it — the evidence gate, the CI split, or the measurements that
changed my mind. Ask and I'll share the application source.
