Skip to content

INSTALL.md

Latest commit

 

History

History
123 lines (87 loc) · 5.49 KB

INSTALL.md

File metadata and controls

123 lines (87 loc) · 5.49 KB
Raw

Installation

This page explains how to install Thundercomm AI Agent on a Thundercomm board, what you need before you start, and how to verify that everything works. The whole install is a single copy-paste command: no .NET runtime, no extra setup, no manual dependency handling.

Prerequisites

Requirement Details
Board Thundercomm TurboX C7790 / TurboX IRB9 (ARM edge AI) or Rubik Pi 3 (12 TOPS, Qualcomm QCS6490)
OS A 64-bit ARM Linux: Debian/Ubuntu-based or Armbian. Other 64-bit ARM distros with systemd and apt work too
Storage At least 1.5 GB free (the release archive is ~700 MB, the installed app ~810 MB)
Memory 4 GB or more recommended
Network Internet access to download the release (GitHub). The assistant itself can later run fully offline with a local model
Access A terminal (SSH or local console) with sudo rights

One-line install

Copy and paste this into the terminal of the Thundercomm:

curl -fsSL https://raw.githubusercontent.com/Graphene-Lab/Thundercomm-AI-Agent/main/install.sh | bash

What the installer does automatically:

  1. Detects the platform (must be 64-bit ARM Linux) and installs curl, tar and libicu if they are missing.
  2. Downloads the latest AgentBridge linux-arm64 release from GitHub (pin a version with THUNDERCOMM_AGENT_VERSION=v1.26.08.30).
  3. Unpacks it directly into /opt/agentbridge and fixes the file ownership.
  4. Creates and starts a systemd service (agentbridge) that keeps the assistant always on and starts it at boot.
  5. Prints the final status.

Options

# Install a specific version instead of latest
curl -fsSL https://raw.githubusercontent.com/Graphene-Lab/Thundercomm-AI-Agent/main/install.sh | \
  THUNDERCOMM_AGENT_VERSION=v1.26.08.30 bash

# Install into a custom folder (no systemd service)
curl -fsSL https://raw.githubusercontent.com/Graphene-Lab/Thundercomm-AI-Agent/main/install.sh | \
  THUNDERCOMM_AGENT_HOME=/home/pi/agent THUNDERCOMM_AGENT_NO_SERVICE=1 bash

First start

After the install the service is already running. Verify it:

curl -s http://localhost:5290/health
# -> {"status":"healthy",...}

curl -s http://localhost:5290/v1/models
# -> lists the agents and the configured AI providers

Terminal chat (TUI): the service owns port 5290, so for the interactive chat run:

sudo systemctl stop agentbridge
/opt/agentbridge/agent

When you are done, put the assistant back always-on:

sudo systemctl start agentbridge

API access: the assistant answers any OpenAI-compatible call on http://<thundercomm-ip>:5290 (e.g. /v1/chat/completions, /v1/audio/speech) — scripts, bots and the web client all use the same server.

Choosing the AI that powers the device

Type /modelsetup in the chat and pick a provider:

  • Local model (fully offline): Ollama or ExLlamaV2 on the board itself.
  • Cloud provider: DeepSeek, Gemini, Anthropic — with the built-in GDPR-ready anonymisation that strips names and identifiers before any request leaves the device.

You can also edit providers.json next to the executable directly (it is the single source of truth for API keys) — see the AgentBridge documentation.

Speech: what works out of the box

  • Text-to-speech (TTS) — yes, verified on ARM: the Kokoro neural voice, its voices and the ONNX model are shipped in the archive. Test it:

    curl -s -X POST http://localhost:5290/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"model":"kokoro","input":"Hello from the Thundercomm.","voice":"af_heart"}' \
  -o /tmp/hello.wav && file /tmp/hello.wav
# -> RIFF ... WAVE audio, 16 bit, mono 24000 Hz

  • Speech-to-text (dictation / voice calls) — the STT engine (voiceagent-stt) is built for Windows and is not yet shipped for Linux ARM. The /v1/voice/listen endpoint reports itself unavailable on Linux until an ARM build of the STT component is published. Text chat, documents, email, research, Telegram, scheduling and podcasts all work on the Thundercomm.

Updating

The assistant updates itself automatically in the background (see the Updates guide). To force a fresh install of a newer release:

sudo systemctl stop agentbridge
curl -fsSL https://raw.githubusercontent.com/Graphene-Lab/Thundercomm-AI-Agent/main/install.sh | bash

Your documents, keys and settings are never touched.

Uninstalling

sudo systemctl stop agentbridge
sudo systemctl disable agentbridge
sudo rm -f /etc/systemd/system/agentbridge.service
sudo systemctl daemon-reload
sudo rm -rf /opt/agentbridge

Troubleshooting

Problem Fix
unsupported architecture The installer only runs on 64-bit ARM (aarch64) Linux — check with uname -m
Service does not start systemctl status agentbridge and journalctl -u agentbridge -n 30 for the error; common causes: missing libicu, or an edited config file with a syntax error
Port 5290 already in use Stop the other instance (sudo systemctl stop agentbridge), or change "Urls" in /opt/agentbridge/appsettings.json and restart
Chat answers fail with "API key is not set" No provider is configured yet — run /modelsetup in the chat, or add the provider in providers.json
Chat answers fail with an error about a local bridge Providers on the same LAN are supported since AgentBridge v1.26.08.30 (they need no API key); point providers.json at the machine that runs the bridge (e.g. http://192.168.x.x:8787/)