Run an offline car assistant on a Raspberry Pi
A local Qwen model on a Pi reads car OBD data and the manual to answer car questions, controls AC and locks via cloud, and alerts other agents to a breakdown so they can look up train tickets.
Done withQwen
- What they did
- A Raspberry Pi 5 (16 GB) rides in the car and runs a local Qwen3.6-35B model through llama.cpp. It reads live OBD data over a Bluetooth ELM327 adapter and answers voice questions from the 489-page owner's manual, with page citations. It posts to a chat room as an agent. The repo says cloud lock and window commands work through Home Assistant, and that precondition (AC) is only the next planned command.
- How it went
- The model runs at about 3.5 tokens/s generation, and the voice, manual and OBD loops are marked proven on a real car. The README does not mention train tickets. Preconditioning and agent-to-agent cases are listed as enabled, not demonstrated.
- Worth knowing
- Loading the 14 GB model from microSD takes about three minutes, and generation is slow. Smaller models run faster on the Pi but suit simple questions better.
Try it yourself with Hermes Agent
Help me build an offline car assistant on a [Raspberry Pi model] using a local model. It should read OBD data from my [car make/model/year] through a [OBD adapter type] and answer questions using the owner's manual PDF at [path]. Finish with a working script I can run and a test question answered from live or sample data; pause before sending any commands that change car settings, such as locks or AC.
Discussion on HN · Aug 25, 2026
Five of these in your inbox every morning
The best things people got an AI agent to do, each with the prompt to try it.
More like this
Get a second opinion on an MRI
One person's account of using Claude Code to analyze their own MRI.
Search LinkedIn and submit tailored job applications
Claude Code drives a real Chrome browser to find matching jobs from a resume, collect them in a CSV, write cover letters, fill forms and submit.
Clear leftover installers and fetch a YouTube transcript
Scout ran PowerShell to delete 31 leftover installers (4.4 GB) after asking permission, and installed a tool to save a YouTube transcript when the browser was blocked.
Monitor and respond to home office sensors
Omar Shahine ran a local OpenClaw agent on his MacBook, extended with home automation so it could observe and respond to physical sensors in his home office.
- Custom agent
Fill out a long GP form
An agent filled out a long GP form, which led to a doctor call and tests being booked automatically.
- Custom agent
Book a vet appointment through a clinic's AI phone bot
A commenter called their doctor about a sick pet and the AI booked an appointment 30 minutes later, which worked out.