Hermes Dashboard (dashu) — a self-hosted workspace
- Type
- Personal project
- Date
- November 2026
- Status
- v1 in service, v2 in progress
- Technologies
- Rust, htmx, SQLite, ffmpeg, Proxmox, LXC, Debian, systemd, Cloudflare, Tailscale, OpenID Connect, GitHub Actions, Lighthouse, Claude Code
- Skills
- B1.1 Manage IT assets; B1.4 Work in project mode; B1.5 Make an IT service available to users; B1.6 Organise one’s professional development
Context
Between my BTS (timetable on Pronote Campus), my work-study job at 1Life and my personal life, my notes, tasks and calendars were scattered across several apps. I wanted one place of my own, hosted at home: a kind of personal Microsoft 365, built to measure rather than an off-the-shelf dashboard.
I also turned it into a project-management exercise: write a real specification with testable acceptance criteria, have it built by an AI (Claude), and keep control of the choices, the deployment and the validation.
Since 12 November 2026, the app is displayed as Hermes Dashboard, like my other projects; dashu remains its code name (repository, program, commands).
Objectives
- A single way in: username and password, or a Google account. Sign-up is closed: only authorised addresses get in.
- Spaces, Personal, School and Work to start with: every note, task, file or calendar belongs to a space, and the Personal space brings together the content of all spaces.
- Markdown notes with search, today's tasks, and the timetable imported from Pronote Campus.
- In version 2, a place for my files (classes, scans, iPhone photos), sorted into folders, with conversions that never put the server at risk.
- Fast (a mobile Lighthouse score of at least 95) and usable on a phone like an app.
- Hosted on my homelab, reachable from anywhere without opening a port on the home router.
- Easy to run: a nightly backup, checked by an actual restore, and an update in one command with an automatic rollback.
Implementation
The specification. I wrote the specification: scope of version 1 and what it leaves out, technologies, data model, performance and security requirements, deployment, and 14 acceptance criteria. A tracking file states, for each criterion, whether it is verified, where and how, and lists the deviations from the specification with their reasons. Every later change (files, spaces and tools, command bar, folders) got its own section, with its own criteria, written before the code.
The application. A single program written in Rust, which builds the pages on the server; the htmx library keeps navigation smooth, without reloading the page. The data lives in an SQLite database, with full-text search. Notes are edited already formatted, as in Notion (a "/" menu inserts headings, lists, tables or images), but are still saved as Markdown. The timetable is synced every hour from the Pronote Campus iCal link. A command bar, opened with "/", makes everything doable from the keyboard. A JSON API, protected by a token, lets Hermès, my AI agent, manage my tasks and notes, and read my calendar.
Security.
- Passwords hashed with Argon2id, and at most 5 login attempts per minute from the same IP address.
- Google sign-in with OpenID Connect, using an OAuth client I created in the Google Cloud Console. An address that is not on the authorised list is refused.
- A strict Content Security Policy (CSP): no inline script or style. Note content is cleaned to prevent script injection.
- The service runs as an unprivileged system user, only listens on the machine itself (127.0.0.1) and sees the file system as read-only (systemd hardening).
- Hermès only has SSH access limited to six actions (status, logs, restart, update, update tracking, backup), enforced by a forced command.
Version 2: files and folders.
- I wiped the server's second disk (465 GB, an old Proxmox install) and declared it as LVM-thin storage in Proxmox. Two volumes from it are mounted in the container: 45 GB for the files, 5 GB for the database backups, which are therefore no longer on the same disk as the database. A third volume, on the first disk this time, receives a copy of the files every night: every piece of data exists on two disks. The copy is incremental (a file already backed up is not copied again, it is hard-linked to the previous copy with rsync), kept 14 days, and restoring one night's database also brings back that night's files.
- A file is sent in 50 MB chunks, because Cloudflare's free plan refuses requests over 100 MB. On disk, it has a random name, never the one chosen by the user. It is never run inside the page: only types without code (images, videos, sounds) are displayed, everything else is served as a download, with a "sandbox" CSP.
- Conversions (iPhone HEIC photos, videos, sounds, PDFs, and Markdown documents with pandoc and Typst) run in a separate service: another system user, with no network, that sees neither the database nor the stored files, limited to 1.5 GB of memory. The web service, limited to 256 MB, never runs a tool itself.
- Folders and subfolders, three levels at most, hold both files and notes. They only exist in the database: nothing moves on disk, and a folder name can never point to another file. Items are filed through a window that shows the whole tree, or by dragging with the mouse; deleting a folder that is not empty goes through a warning.
- Files have the same tags as notes, and a search on their name (SQLite full-text index, accents ignored) in the open folder and its subfolders. A file can be renamed from its preview or from the list: only the part before the extension changes, the file type stays the one it was uploaded with.
- A file can be shared through a link valid for 1 hour, 1 day or 7 days, and revocable. The link holds 32 random bytes; the database only keeps their SHA-256 hash, so reading it does not give a link back. The link's page only shows the name and the size, and downloads the file; expired, revoked or unknown, it is the same page, and the number of requests per IP address is limited.
Deployment on my homelab.
- A Debian 13 LXC container on Proxmox. dashu is compiled there and runs as a systemd service.
- A Cloudflare tunnel: the container opens an outgoing connection to Cloudflare, which publishes app.dashu.dev over HTTPS. No port is open on the home router. I moved the DNS of the dashu.dev domain to Cloudflare.
- Tailscale: a second way in, through my VPN, which works even if Cloudflare is down.
- dashu.dev shows a public page that presents the project (GitHub Pages).
Operations.
- A hot backup of the database every night at 3:30 a.m. (systemd timer), checked by SQLite and kept 14 days, then the copy of the files to the other disk.
- dashu-restore puts a backup back into service: it first sets the current database aside, and rolls back by itself if dashu does not restart.
- dashu-update fetches the code, recompiles, backs up the database, replaces the program and checks that the service responds. If not, it puts the previous version back.
- On every change, GitHub Actions checks the formatting and quality of the code, runs the automated tests (247 today), replays the server's systemd sandbox, tests the file backup script and builds a Docker image under 30 MB.
Project tracking. After version 1, every change went through a pull request, tested before being merged: the API for Hermès, an overview of all calendars in the Personal space, navigation without a black flash, a "Dynamic Island" style space switcher with swipe-to-switch, a shortcut bar on phones, installation as an app on the iPhone home screen. Then custom spaces (up to eight), a modular dashboard, version 2 with files, the command bar that replaces the Ctrl+K palette, the new note editor and folders.
Built with Claude Code. My role: define the need, make the choices and test. I did the deployment and operations myself on my server (Proxmox, Cloudflare, Tailscale, Google, backup and restore), guided step by step.
Source code: github.com/moreauremi/dashu · Project page: dashu.dev
Screenshots
Screenshots of version 1, from 11 November 2026, taken on a copy of dashu filled with demo data (fake classes, fake notes). The last one is the real output of the restore on my server. Since then, the editor shows a single formatted view, and the Ctrl+K palette has become the command bar.






Results and testing
- All 14 acceptance criteria of the specification are verified, most of them on the real server.
- 96 automated tests at the end of version 1, 247 today (sign-in, data isolation between two accounts, notes, reading the timetable including daylight saving time, API, script injection, booby-trapped files, folders, expired or revoked share links), run by GitHub Actions on every pull request.
- Mobile Lighthouse on the server (version 1), through Cloudflare, on simulated 4G: 99 for the dashboard, 100 for the notes, 100 for the calendar, 99 for the editor.
- At the end of version 1, the service used 7.6 MB of memory and responded in 0.5 ms; the Docker image was 7 MB.
- Google sign-in tested with my account (the same as with my password) and with an unauthorised account, which is refused.
- Actual restore on 11 November 2026: a note created after the backup disappears when it is restored, then comes back when the restore is undone.
- Compression checked on the server, through the isolated service: a 4032 × 3024 HEIC photo becomes a 2560 × 1920 JPEG (816 KB → 391 KB), a PDF goes from 626 KB to 57 KB, a 10-second 4K HDR video becomes a 3.7 MB 1080p MP4 instead of 54.8 MB. During a conversion, the service only sees the local network interface.
- Folders tried on the server on 15 November 2026, from my iPhone and my computer.
- I have used Hermes Dashboard every day since November 2026, on my computer and my iPhone.
Difficulties encountered
The site unreachable on my home Wi-Fi. app.dashu.dev worked on 4G but not at home. The router's DNS was still holding an old answer, "this name does not exist", from before the domain moved to Cloudflare. I saw it by comparing the answers of the router's DNS and of Cloudflare's (1.1.1.1). On a Mac, nslookup queries the DNS server directly, whereas the browser goes through the system cache: dscacheutil showed me what the browser actually saw. Solution: my devices use the DNS servers set by Tailscale instead of the router's.
An update that could be wrong. Reviewing the update script revealed a flaw: if compilation failed, the source code was already up to date but the running program was not, and the next update replied "already up to date" without recompiling. Fix: the version actually in service is recorded separately, and the status command reports any mismatch.
A backup that has never been restored proves nothing. The backup ran every night, but nothing showed it could be put back. I did the drill on the server without losing anything: the current database is set aside before each restore, and a test note proves that the data really comes from the backup.
Zoom on iPhone. Added to the home screen, the app could be zoomed like a web page, and iOS has ignored the standard setting since iOS 10. Solution: a web app manifest, and pinch blocking only in app mode. In Safari, zoom is still possible, for accessibility.
Compression that worked everywhere except on the server. The "Compress" button failed in production while every test passed. Two causes, specific to the service's systemd sandbox, which is absent locally and in the tests: RestrictSUIDSGID=yes refuses any chmod that sets the setgid bit, and each folder listed in ReadWritePaths is a separate mount point, so a plain rename from one folder to the other fails (EXDEV). Fix: no more setgid bit, and a copy followed by a delete when the rename fails. So that the bug does not come back, the CI now replays this sandbox.
A skewed measurement. The first Lighthouse run on the dashboard scored about 94, below the target: my browser extensions added work while the page loaded. I measured again three times in a Guest window, with no extensions: median of 99.
Possible improvements
- Run the restore drill with a file on the server: delete one, then bring it back from the nightly backup.
- Monitor the service with Uptime Kuma, in its own container (the procedure is written), to be told when it goes down.
- Finish version 2: opening files to Hermès's API.
Skills applied
- B1.1: manage backups (checked nightly backup, actual restore, cross backups: the database on the second disk, the files on the first); set up and check access levels (authorised addresses, API token, Hermès's SSH access limited to six actions, temporary and revocable share links); check service continuity (two ways in, automatic rollback of updates, /health check endpoint).
- B1.4: write a specification with acceptance criteria, split the work into pull requests, track progress criterion by criterion and analyse deviations from the specification.
- B1.5: carry out integration and acceptance testing (automated tests, criteria checked on the server, Lighthouse measurements), deploy the service (Proxmox, LXC, LVM-thin storage and mount points, systemd, isolated conversion service, Cloudflare, Tailscale, DNS, Google OAuth) and document how to install and run it.
- B1.6: set up my personal work environment, with my classes, revision notes and timetable in one place, and use it every day.