Your car as a chat-room agent — fully offline. A Raspberry Pi 5 rides in the car, runs a 35B-parameter model locally, joins your GroupMind rooms as @gle (or whatever you name yours), and messages you like any other agent: departures, arrivals, trip summaries, and dashcam clips when something hits the car — with approvals and replies from your phone or watch via CodeWatch. Open PRs land on the CodeWatch dashboard next to ClawWatch and WhereWatch.

Live and measured, on real hardware (Pi 5, 16 GB, ~300 €):

• 🧠 Qwen3.6-35B-A3B (Unsloth UD-Q3_K_S dynamic quant, 14.3 GB) at 3.5 tok/s generation / 25+ tok/s prompt, 65 °C sustained, no cloud, no internet, no subscription.
• 📖 Answers from the car's own 745-page owner's manual with page citations (lexical RAG, ships on the SD card) — and refuses to answer what the manual doesn't say.
• 🔬 Grounded self-knowledge: temperature, throttling, fan, memory, disk, network and which model is loaded are read live from the machine per question. What it can't sense, it says it can't sense — the system prompt is built so an unknown can never silently read as a fact.
• 🎙️ Hands-free voice: a continuous listener (energy VAD → whisper.cpp, all on-Pi) hears you speak, routes the words through the same grounded pipeline, and answers into the room. No wake word ceremony, no cloud STT.
• 📡 Autonomous: systemd services self-start the whole stack on boot — model server, room agent, voice listener, phone dashboard, engine watcher.
• 🔧 Maintainable from anywhere: the car pulls its own updates from this repo (hourly + a dashboard "update now" button) and dials out a tunnel so it stays reachable even behind a phone hotspot's NAT. No laptop-in-the-car maintenance, ever.
• 📶 Three-tier connectivity: phone hotspot → home wifi → its own fallback access point, so the phone can always reach it, even in a garage with zero signal.

The build log with every dead end included lives in docs/plan.md.

Sibling of CodeWatch (agents on your wrist; source: codewatch-cli) and ClawWatch (health on your wrist; v2 launch video). This one watches the car. The rest of the family lives at thinkoff.io.

flowchart LR
    subgraph car [In the car - Raspberry Pi 5]
        MIC[USB mic] --> LISTEN[carwatch-listen<br/>VAD + whisper.cpp]
        LISTEN --> BRAIN[llama.cpp server<br/>Qwen3.6-35B-A3B]
        MANUAL[(Owner manual RAG<br/>745 pages, on SD)] --> BRAIN
        STATE[selfstate<br/>temp / fan / net / model] --> BRAIN
        OBD[carwatch-obd<br/>watches the OBD cable] --> AGENT
        BRAIN --> AGENT[carwatch-agent<br/>the @gle room agent]
        DASH[web dashboard :8088<br/>status / update / voice / wifi]
        UPD[self-update<br/>hourly git pull] -.updates.-> car
        REACH[dial-out tunnel<br/>reachable behind any NAT]
    end
    AGENT <-->|posts + mentions| GM[GroupMind rooms]
    GM <--> PHONE[Your phone / watch<br/>CodeWatch]
    DASH <-->|same wifi| PHONE

Local is the product; online is the enrichment. The car must be fully useful with zero connectivity, because cars live in garages, tunnels and countryside dead zones:

• Always local (works with no signal): voice in, the assistant's answers (on-Pi model), owner's-manual answers (RAG ships on the SD card), the phone dashboard (served BY the car), trip/state tracking.
• Queued through connectivity gaps: room posts, clip uploads, mention replies. Everything lands in a persistent on-disk outbox first and is delivered late rather than lost.
• Online-only, and honest about it: remote reachability (the dial-out tunnel), self-updates, escalation to bigger brains — first a local-LAN model server when one rides along (still no cloud), then a cloud model only when online AND explicitly asked, on the car's own budget-capped key.

Rule of thumb: glanceable safety-relevant info never depends on the network; anything social or heavy degrades gracefully to "later".

A car keeps four palm-sized contact patches on the road, the only place it ever meets reality. One principle per wheel: assert only what you can sense, claim only what is verified, label anything interim loudly, and report failure plainly with no silver lining. Everything above those four patches is just suspension.

— @claudeMB, CarWatch dev log, after a day of learning all four the hard way

Honesty policy: a feature is only "proven" after it worked on the real car. "Built + tested" means the code runs end-to-end against a real or simulated counterpart but has not yet met the physical car.

• Raspberry Pi 5, 16 GB (active cooling required — the SoC throttles without it)
• USB microphone for voice (any class-compliant mic)
• WOLFBOX G900 3-channel dashcam (wifi AP; CarWatch pulls event clips from it)
• OBD access: ethernet-to-OBD (DoIP/ENET) cable — support built, real-car verification pending; a standard ELM327-class adapter is the fallback path
• Power: the dashcam hardwire kit feeds the camera; the Pi needs its own 5V/5A USB-C feed (12V PD adapter, or the car's 230V socket + wall PSU)

git clone https://github.com/ThinkOffApp/CarWatch.git
cd CarWatch
./install.sh

Then put your credentials in /etc/carwatch/config.json (never in the repo — see config.example.json) and:

sudo systemctl enable --now carwatch

After that the car keeps itself current: update.sh pulls this repo's main, installs any new systemd units, and restarts services — on a timer, from the dashboard button, or by hand:

curl -sSL https://raw.githubusercontent.com/ThinkOffApp/CarWatch/main/update.sh | bash

Copy config.example.json to /etc/carwatch/config.json:

• api_base — your GroupMind server, e.g. https://groupmind.one
• api_key — the agent's API key (create one for the car; never reuse another agent's key, never commit it)
• room — room slug the car posts to
• handle — the car's display handle, e.g. @gle
• home_ssids — wifi networks that mean "parked at home"
• wolfbox — dashcam AP name/password and poll interval

The WOLFBOX's HTTP API is undocumented; carwatch-probe discovers it:

python3 -m carwatch.wolfbox --probe

Connect the Pi to the dashcam's wifi AP first. The probe walks known dashcam-firmware endpoint patterns and prints what answers, which fills in wolfbox.py's TODOs with your camera's real paths.

AGPL-3.0, like ClawWatch. Copyright (C) 2026 ThinkOff / Petrus Pennanen.