Private, on-device, voice-first civic concierge for London.
Ask in plain language - "where do I vote and is it step-free?", "nearest accessible toilet to Brixton station?" - and get a spoken answer grounded in real council open data, on a device that keeps your questions and location to itself.
- Surfaces civic data commercial maps don't carry - polling stations, emergency rest centres, grit bins, council CCTV coverage, accessible-toilet detail - across 8 London authorities, by asking in plain language via voice or text.
- Routes safely - walking routes prefer monitored, well-served streets; CCTV and grit-bin density are scored on-device from the same GeoJSON, not from a remote service.
- Stays on the box - reasoning, civic lookups, postcode geocoding, route-safety scoring and conversation memory never leave the device. Voice (ElevenLabs) is the only network call, and it can be replaced with on-device Whisper.
Built for the people commercial tools underserve: new arrivals, the elderly, the visually impaired, and anyone who needs civic access rather than the nearest café.
Validate the data engine - no install, no GPU, no API key:
node plugins/civic-geo/scripts/smoke.mjsExpected: Triton Square → Regent's Park library at ~415 m; 41 cameras within 500 m of Brixton. If it looks right, the civic data logic is proven independently of the LLM or the box.
Deploy the full stack (requires an NVIDIA GB10 box and an ElevenLabs API key):
cp .env.example .env && $EDITOR .env # set ELEVENLABS_API_KEY - the only required secret
# GB10 shares one 128 GB unified pool between CPU and GPU; free it before loading the model:
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'
free -h # want `available` comfortably above ~45 GiB
make gate-test # confirm Nemotron NVFP4 serves on the GB10 (~16 GB weights, downloaded once)
make demo # stage weights → build → docker compose up --wait → pre-warmThen open http://<box>:8091 in a browser.
SSH tunnel (from a laptop with no LAN access):
ssh -N -L 8091:127.0.0.1:8091 <user>@<box>
# then open http://localhost:8091Quick text test without the UI:
docker compose exec gateway openclaw agent --agent main --message "nearest library to 1 Triton Square"Full runbook (topology, GB10 memory gotchas, integration points to verify, raw Compose commands) →
deploy/DOCKER.md. Bare-metal path (no Docker) →
docs/setup-runbook.md.
Four Docker containers on one GB10 box, connected by two Docker networks that enforce the privacy boundary at the infrastructure level.
Full-resolution interactive diagram → docs/architecture.html
(htmlpreview link
when the repo is public - GitHub can't render HTML directly.)
Spine (never leaves the box): browser → bridge :8091 → gateway (OpenClaw + civic-geo
plugin) → vllm (Nemotron reasons, calls tools) → answer back through the same path.
Voice excursion (spoken turns only): ① browser mic → ② bridge sends audio out through
egress-proxy to ElevenLabs Scribe (STT) → ③ spine runs → ④ bridge sends answer text out to
ElevenLabs v3 (TTS) → spoken reply. A typed turn uses only the spine and never touches
ElevenLabs or any network.
The egress-proxy is a default-deny allowlist: *.elevenlabs.io only, everything else 403. The
vllm and gateway containers sit on an internal: true Docker network with no gateway/NAT -
they cannot initiate outbound connections. The privacy boundary is enforced by Docker, not just
asserted.
Always on the box: reasoning, civic-data lookups, postcode geocoding, route-safety scoring, conversation memory. What you asked and where you are never leaves the device.
Three optional network calls, each independently disableable:
| Call | Purpose | How to disable |
|---|---|---|
| ElevenLabs voice in/out | STT (Scribe) + TTS (v3) | Set LOCAL_STT_CMD for on-device Whisper STT; mute TTS to go fully offline |
| Walking-route geometry (OSRM) | Turn-by-turn step directions | ROUTING_DISABLE=1 - route-safety scoring stays on-device either way |
| Assigned polling station (gov API) | Your assigned station, not just nearest | Unset POLLING_LOOKUP_URL - falls back to nearest facility, honestly labelled |
With all three disabled the system is air-gapped end to end. None sends your identity: OSRM sees two coordinates, the polling API sees a postcode, ElevenLabs sees audio or text.
The components are designed to be extended independently:
| What to extend | How |
|---|---|
| More London boroughs | Add datasets/<facility>/<borough>/<facility>.geojson; the engine picks it up at startup with no code changes |
| More facility types | Add a GeoJSON folder + a category entry in plugins/civic-geo/src/geo.mjs; the OpenClaw layer and LLM need no changes |
| Assigned polling lookup | Set POLLING_LOOKUP_URL to a Democracy Club API template with {postcode} |
| On-device voice (no ElevenLabs) | Set LOCAL_STT_CMD to a Whisper command for STT; TTS is independently mutable |
| Different LLM | Swap agents.defaults.model.primary in deploy/openclaw.gateway.json; civic-geo is model-agnostic |
| Embed + rerank + content-safety | Add extra vLLM services + a LiteLLM router; point the gateway at the router |
| Different city | Replace datasets/ with any city's open GeoJSON; update category mappings in geo.mjs |
The clean separation is: civic-geo owns data logic (pure Node, zero deps, independently testable);
OpenClaw owns agent runtime, voice and memory; the bridge owns the web surface. Each layer can be
swapped without touching the others.
- Dataset coverage: 8 of 33 London authorities. Only those that publish civic facilities as
open location data under OGL; the other 25 have no open location data portal. Full coverage
matrix, sources, caveats and refresh URLs →
datasets/SOURCES.md. - Nearest ≠ assigned. Returns the nearest polling station or school, not your legally assigned
one. Assigned lookup requires a postcode-to-district API; see
POLLING_LOOKUP_URLabove. - Geocoding is approximate. Landmark gazetteer + address/name fallback - accurate enough for
"near Brixton", not OS-grade surveyed points. Add entries to
geo.mjs LANDMARKSfor any place the product will name by landmark. - CCTV = traffic cameras. The CCTV layer is TfL traffic/town cameras (busy, well-served roads),
not community-safety surveillance.
safety_countdocuments this and the agent repeats it. - Hardware-specific. The containerized stack is built and tested for the GB10 (arm64, 128 GB
unified memory, CDI GPU passthrough). Other NVIDIA hardware needs
VLLM_GPU_FRACand--max-model-lentuning; x86 needs a different vLLM image tag.
plugins/civic-geo/ on-device GeoJSON tool plugin - the team's core IP (see its README)
datasets/ London civic facility GeoJSON, per borough + merged all-london files (OGL)
ui/ accessible, answer-first web UI + zero-dep Node bridge (bridge.mjs)
deploy/ containerized stack: Dockerfiles, OpenClaw configs, egress allowlist
docker-compose.yml single-command bring-up
Makefile orchestration (gate-test, demo, prove-boundary, logs, nuke)
docs/ architecture diagram (HTML + PNG), setup runbook, pitch deck
datasets/SOURCES.md coverage matrix, sources, licences, field caveats, refresh URLs
docs/london-structure.md why facility data is split across 33 separate council portals
The highest-impact contribution is additional borough datasets: if your council publishes any of
these facilities as open location data, open a PR adding a GeoJSON file under
datasets/<facility>/<borough>/ - the engine picks it up with no code changes. See
datasets/SOURCES.md for the format, coverage gaps, and which 25 boroughs
are still missing.
For code contributions, plugins/civic-geo/README.md is the right
starting point - that's where the data logic lives and it has a zero-install test harness.
London civic facility data is sourced borough-by-borough from council open data portals, all
released under the UK Open Government Licence (OGL v3).
OGL permits redistribution, reuse and adaptation with attribution. Sources and attribution per
borough → datasets/SOURCES.md.
Codeborough was built at NVIDIA Hack for Impact - London (June 2026, Public Services track) in under 24 hours. The IP belongs to the team; the code is MIT-licensed.