File size: 9,263 Bytes
66fd6b1 4b97c3b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 | ---
language:
- en
- es
tags:
- iot
- smart-home
- home-assistant
- tool-calling
- agent
- cybersecurity
- devops
- ollama
- inference-pipeline
- autonomous-agent
- raspberry-pi
- qwen3-coder
license: mit
library_name: transformers
pipeline_tag: text-generation
---
# TokioAI-Coder IoT Package
**Turn a 30B open-weight model into a fully autonomous IoT/smart-home agent.**
This package contains everything needed to deploy TokioAI-Coder as an autonomous IoT controller:
- The complete **TokioAI Engine** -- inference pipeline that wraps any Ollama model and adds reliable tool calling
- A production-tested **IoT system prompt** with a full Home Assistant cookbook
- **12 native tools** for system administration, file operations, network diagnostics, and robot vision
- **Server-side investigation** -- the engine autonomously chains 8+ tool calls to answer a single question
## What It Does
You say: *"is Home Assistant running?"*
The engine:
1. Checks `systemctl status home-assistant` -- not found
2. Checks `docker ps | grep home` -- not found
3. Checks `ss -tlnp | grep 8123` -- port listening
4. Checks `curl -s http://localhost:8123/api/` -- HA API responds
5. Answers: "Yes, HA is running on port 8123 via Docker"
All 5 steps happen **in a single API call**. The client sees only the final answer.
You say: *"play some blues"*
The engine:
1. Calls HA API to send `play_media` to the Alexa
2. Waits 3 seconds
3. Queries the media player state to verify the music actually changed
4. Reports: "Now playing: Blues Essentials on Jarvis"
## Architecture
```
TokioAI Engine (OpenAI-compatible API)
====================================
User Query ──> Router ──> Ollama Model ──> Verifier ──> Tool Executor ──>
| | | |
Intent System Retry Local
Classifier Prompt + Loop (2x) Execution
(0ms regex) Few-Shot (read-only safe)
| | | |
v v v v
tool_call Guided Validates Feeds result
/ text / Generation JSON + back to model
decision (schema- correct for next step
constrained) tool names
|
v
Investigation Loop
(up to 8 rounds,
90s timeout)
|
v
Final Answer
```
## Components
### 1. TokioAI Engine (`engine/`)
The core inference pipeline (3,774 lines of Python):
| Module | Lines | Purpose |
|---|---|---|
| `server.py` | 864 | OpenAI-compatible API with server-side investigation |
| `schemas.py` | 497 | Tool definitions, JSON validation, parameter schemas |
| `pipeline.py` | 476 | Main orchestrator: Router -> Memory -> Model -> Verifier |
| `router.py` | 382 | Intent classifier (regex) + system prompt builder |
| `memory.py` | 353 | Adaptive few-shot memory (learns from successes) |
| `executor.py` | 331 | Safe server-side tool execution |
| `guided.py` | 228 | JSON schema constraints for Ollama |
| `verifier.py` | 186 | Output validation + retry logic |
### 2. IoT System Prompt (`prompts/iot_system_prompt.md`)
A production-tested system prompt that includes:
- Complete Home Assistant API cookbook (media, lights, switches, sensors, vacuum, TTS)
- Alexa voice control via `alexa_media` integration
- Robot vision via moondream (1.4B vision model)
- Investigation playbooks for multi-step problem solving
- Anti-hallucination rules (never simulate tool execution)
- Verification protocol (always confirm state changes actually happened)
### 3. Tool Definitions (`tools/`)
12 native tools in OpenAI function-calling format:
- `execute_local` -- Run shell commands
- `execute_raspi` / `execute_gcp` / `execute_router` -- Remote execution
- `read_file` / `write_file` / `edit_file` -- File operations
- `search_files` -- grep across codebases
- `diagnose` -- System health checks
- `robot_vision` -- Camera capture + AI analysis
- `memory` -- Persistent memory across sessions
- `task` -- Task tracking for multi-session projects
## Quick Start
### Prerequisites
- Python 3.8+
- [Ollama](https://ollama.ai/) with a model (recommended: `qwen3-coder` 30B MoE)
- Home Assistant instance (for IoT features)
### Install
```bash
# Clone this repo
git clone https://huggingface.co/MrMoz33/tokioai-coder-iot
cd tokioai-coder-iot
# Install dependencies
pip install flask requests
# Pull a model
ollama pull qwen3-coder # 30B MoE (3B active), recommended
# or: ollama pull qwen2.5-coder:7b # lighter alternative
```
### Configure Home Assistant
Set your HA token:
```bash
export HA_TOKEN="your-long-lived-access-token"
export HA_URL="http://localhost:8123" # or your HA URL
```
### Start
```bash
# Start the engine
python -m engine serve --model qwen3-coder --port 8080
# Use it (any OpenAI-compatible client works)
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "tokioai-engine",
"messages": [{"role": "user", "content": "is home assistant running?"}]
}'
```
### With TokioAI CLI
```bash
pip install git+https://github.com/TokioAI/tokioai.git
export TOKIOAI_ENGINE_HOST=http://localhost:8080
tokioai # Interactive agent with full tool calling
```
## Customization
### Adding Your Own Home Assistant Entities
Edit `prompts/iot_system_prompt.md` and add your entities to the cookbook section:
```markdown
LIGHTS: light.living_room, light.bedroom, light.kitchen
SWITCHES: switch.fan, switch.heater
SENSORS: sensor.temperature, sensor.humidity
MEDIA: media_player.living_room_speaker
```
### Adding Custom Tools
Edit `engine/schemas.py` to add new tool definitions:
```python
{
"type": "function",
"function": {
"name": "my_custom_tool",
"description": "What this tool does",
"parameters": {
"type": "object",
"properties": {
"param1": {"type": "string", "description": "First param"}
},
"required": ["param1"]
}
}
}
```
Then add the executor in `engine/executor.py`.
### Adapting for Different Smart Home Platforms
The system prompt is platform-agnostic in concept. To use with:
- **Google Home**: Replace HA API calls with Google Home API
- **Apple HomeKit**: Use HomeKit Accessory Protocol
- **MQTT Direct**: Replace curl commands with `mosquitto_pub/sub`
The engine and tool pipeline remain identical -- only the system prompt changes.
## Performance
Tested on Raspberry Pi 5 (8GB) + L4 GPU (via SSH tunnel):
| Metric | Value |
|---|---|
| Model | Qwen3-Coder 30B (3B active MoE) |
| GPU VRAM | 19.4 GB (L4 24GB) |
| Response time | 3-8s (simple), 15-45s (investigation) |
| Investigation depth | Up to 8 rounds per query |
| Tool calling accuracy | 94%+ with engine pipeline |
| Uptime | 30+ days continuous (systemd service) |
### Cost Comparison
| Approach | Monthly Cost | Latency | Privacy |
|---|---|---|---|
| GPT-4o API | $50-200 | 2-5s | Data sent to cloud |
| Claude API | $50-200 | 2-5s | Data sent to cloud |
| **TokioAI-Coder (self-hosted)** | **$0** | **3-8s** | **100% local** |
After initial GPU investment ($0.50-1.50/hr for L4 on GCP, or free on local hardware), all inference is free and private.
## Model Compatibility
| Model | Params | Tool Calling | IoT Quality | Notes |
|---|---|---|---|---|
| qwen3-coder | 30B (3B active) | Excellent | Excellent | Recommended. MoE = fast inference |
| qwen2.5-coder:7b | 7B | Good | Good | Budget option |
| llama3.1:8b | 8B | Good | Fair | General purpose |
| deepseek-coder-v2:16b | 16B | Good | Good | Strong code understanding |
| mistral:7b | 7B | OK | Fair | Needs guided gen |
The engine's guided generation + retry logic compensates for weaker models.
## Related Projects
- [TokioAI CLI](https://github.com/TokioAI/tokioai) -- Multi-provider AI CLI with native tool calling
- [TokioAI Fine-Tuning Guide](https://github.com/daletoniris/tokioai-fine-tuning-guide) -- Train your own tool-calling model
- [TokioAI Fine-Tuning Paper](https://github.com/daletoniris/tokioai-fine-tuning-paper) -- Research paper on LLM fine-tuning for tool use
- [TokioNav](https://github.com/daletoniris/tokionav) -- AI-controlled autonomous robot navigation
## License
MIT
## Credits
Built by Daniel Dieser ([daletoniris](https://github.com/daletoniris)) and the TokioAI team.
The TokioAI Engine was developed to make small open models competitive with GPT-4 and Claude for autonomous tool calling -- specifically for IoT, cybersecurity, and DevOps tasks. The key insight: you don't need a bigger model, you need a smarter pipeline around it.
|