Copy of github.com/askrobots/network-computer (MIT): a desk you reach from anything. The original lives on GitHub.
One VM, everything: voice AI, the dbbasic apps, and the object server
Plan, 2026-09-22. Where the network computer goes next: a VM you talk to, running the dbbasic apps on its desktop and the object server as its brain.
Where we are
https://nc.dbbasic.com streams a desktop with audio out, microphone in (phone app; browsers and Safari now that it has a real certificate), keyboard/mouse/touch, token auth, pairing, and per-session relay credentials. provision/verify.sh checks 34 things on every provision. So the box can already hear you and speak to you; nothing on it listens or answers yet.
1. One TLS front door (Caddy)
Today the rendezvous owns ports 80/443 for its Let's Encrypt cert. The object server's own install path puts Caddy (Apache-2.0) on 80/443 in front of 127.0.0.1:8001. Both can't hold the ports, and more apps are coming, so Caddy becomes the one front door:
| Hostname | Caddy proxies to | What |
|---|---|---|
nc.dbbasic.com | 127.0.0.1:8765 (WebSockets too) | rendezvous + web client |
objects.<test-domain> | 127.0.0.1:8001 | object server |
| later: one per app | its loopback port | other services |
The rendezvous then listens on loopback only and needs a small -behind-proxy flag so /config reports secure (trusting X-Forwarded-Proto from 127.0.0.1 only). STUN/TURN stays on 3478/udp directly; Caddy doesn't need to see it.
Careful: dbbasic.com itself is served in production by an object server elsewhere. The test VM must use a separate hostname so nothing here touches the live site.
2. The object server (optional profile)
askrobots/dbbasic-object-server: stdlib Python + uvicorn, live versioned objects, an app suite (projects, notes, tasks, contacts, articles, links, calendar, files, a shell you can talk to), one permission engine across web/API/search/files/MCP/AI. Its scripts/install.sh is idempotent and makes a systemd service on 127.0.0.1:8001 with an admin token in /etc/dbbasic-object-server.env.
provision/profiles/object-server.sh: clone, run its installer, add the Caddy site.- Profiles are opt-in:
NC_PROFILES=object-server,appsin/etc/nc/env, recorded
like NC_HOST_NAME so a re-provision keeps them.
verify.shgains/health= 200 and the Caddy route, only when the profile is on.
3. The dbbasic apps on the desktop
Inventory of /Volumes/T9 (2026-09-22):
| Kind | Projects | On the VM as |
|---|---|---|
| Flutter desktop apps | dbbasicshell (terminal), dove (email), writer-native, spreadsheet-native, slides-native, draw, slate, scroll, webmaster, porter | Linux bundles in /opt/dbbasic/<app> + menu/panel launchers |
| Mobile-only | tape (recorder) | skip on desktop |
| Hub | object server | service behind Caddy (profile) |
| Python services / tools | content, forms, passwd, census, cli, video, website, studio, tsv (library), dbbasic | services where they serve, libraries where they don't |
| Node | dbbasic-browser (x402) | desktop app once packaged |
| Unclear / empty | 1brc, calc, lens, perf, player | check before planning |
Delivery reuses the build box (docs/BUILDBOX.md): build Linux bundles, then provision/profiles/apps.sh unpacks them and writes .desktop entries. Each project's DBBASIC.md gains install: and desktop: next to its existing build:, so one sweep builds, installs on the VM, and publishes to dbbasic.com.
4. Talking back and forth with an AI
The audio plumbing exists; this adds a listener and a voice:
phone mic → nc-mic-in (done) → end-of-turn → speech-to-text → AI → text-to-speech → nc sink → phone (done)
Open-source parts: whisper.cpp (MIT) for speech-to-text, Piper (MIT) for the voice, Silero VAD (MIT) for end-of-turn later. The AI is either the object server's AI shell (so it can act on your notes/tasks through the same permission engine) or a Claude API key held on the VM, never on the client.
Phases, each shippable:
- Push-to-talk. Hold the mic button; release ends the turn. No detection errors.
Transcript and reply appear in a desktop window as well as out loud.
- Hands-free. VAD decides end-of-turn. The phone's echo cancellation already keeps
the AI's own voice out of the microphone.
- Barge-in. Speaking stops the AI mid-sentence.
- Actions. It can do things: object server data over its API/MCP, and the desktop
itself (shell + GUI, docs/AGENT.md), with a spoken confirmation before anything destructive.
Latency target: under ~1.5 s from end of speech to first sound. Rough budget: speech-to- text 0.3–0.7 s (whisper small.en on 4 vCPU), first AI tokens 0.3–0.8 s, first spoken sentence 0.1–0.2 s (Piper, streamed per sentence), network ~0.05 s.
Safety: push-to-talk first, the mic button is already red while live, the AI runs as its own user with least privilege, and its key lives in the VM's secret store.
5. Working together
- The object server is the shared memory: notes, tasks, contacts, files, permissions.
- Desktop apps read and write it over its HTTP API (dove turns mail into tasks, writer
saves articles, draw saves files).
- The AI sees both sides: data through MCP, the screen through the network computer.
DBBASIC.mdin every project ties build, install, desktop entry and publishing together.
Desks
Storage by lifecycle, hot desking, and automatic start and stop are planned in DESKS.md. Models live on a separate models volume there.
Sizing
Desktop + object server + whisper wants more than today's 2 vCPU / 4 GB: s-4vcpu-8gb (~$0.07/hr, ~$48/month). Speech-to-text can move to a GPU box later.
Order of work
- Caddy front door; rendezvous behind it (
-behind-proxy). verify.sh stays green. - Object server profile on its own test hostname; health check in verify.sh.
- Push-to-talk voice loop with a transcript window.
- Linux builds of three flagship apps (dbbasicshell, writer, dove) as desktop apps.
- Hands-free and barge-in; AI actions with confirmations.
- The remaining apps;
install:/desktop:in DBBASIC.md; one sweep for everything.
Decisions needed
- Test hostname for the object server (must not be
dbbasic.comitself). - The AI: object server's AI shell, a Claude API key, or both.
- Whether to move the test VM to 4 vCPU / 8 GB before the voice work.