- Python 99%
- Shell 0.5%
- Dockerfile 0.5%
Updated channel utilization scoring thresholds: - UTIL_HEALTHY: 15% -> 20% (channel is clear) - UTIL_CAUTION: 20% -> 25% (slight degradation) - UTIL_WARNING: 25% -> 35% (severe degradation) - UTIL_UNHEALTHY: 35% -> 45% (mesh struggling) Previous thresholds were overly conservative. New values better reflect actual Meshtastic firmware behavior and when operators should take action. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> |
||
|---|---|---|
| .github/workflows | ||
| dashboard-frontend | ||
| meshai | ||
| .dockerignore | ||
| .gitignore | ||
| config.example.yaml | ||
| docker-compose.yml | ||
| docker-entrypoint.sh | ||
| Dockerfile | ||
| LICENSE | ||
| pyproject.toml | ||
| README.md | ||
| requirements.txt | ||
MeshAI
LLM-powered mesh intelligence assistant for Meshtastic networks. MeshAI connects to your mesh as a physical node, monitors network health in real-time, and answers questions about your infrastructure over LoRa.
What It Does
MeshAI runs on your Meshtastic node and provides:
- Mesh Intelligence — 5-pillar health scoring, per-region breakdowns, infrastructure monitoring, coverage gap analysis, and environmental sensing
- Conversational Queries — ask "how's the mesh?" or "tell me about MHR" and get data-driven answers over LoRa
- Node Distance — GPS-based distance calculations between any two nodes on the mesh
- Multi-Source Awareness — aggregates data from multiple Meshview instances and MeshMonitor with staggered polling
- Feeder Gateway Tracking — identifies which physical MQTT gateways hear each node and signal quality
- Subscriptions — scheduled daily/weekly health reports and instant alerts delivered via DM
- LLM Chat — general conversation, knowledge base lookups, and weather queries
- Multi-Backend — supports Google Gemini, OpenAI, Anthropic Claude, and local LLMs via LiteLLM
Quick Start
# Clone
git clone https://github.com/zvx-echo6/meshai.git
cd meshai
# Install
pip install -e .
# Configure (interactive TUI)
meshai --config
# Run
meshai
Or with Docker:
mkdir -p meshai/data && cd meshai
curl -O https://raw.githubusercontent.com/zvx-echo6/meshai/main/docker-compose.yml
curl -o data/config.yaml https://raw.githubusercontent.com/zvx-echo6/meshai/main/config.example.yaml
# Edit data/config.yaml
docker compose up -d
Commands
| Command | Description |
|---|---|
!health |
Mesh health overview with colored status dots |
!region |
List all regions with health status |
!region [name] |
Detailed region breakdown |
!neighbors [node] |
Top infrastructure neighbors with signal quality |
!sub daily 6pm |
Subscribe to daily health reports |
!sub weekly 8am sun |
Subscribe to weekly digest |
!sub alerts |
Subscribe to instant alerts on issues |
!unsub [type] |
Remove a subscription |
!mysubs |
List your active subscriptions |
!clear |
Clear conversation history |
!help |
Show available commands |
!help [cmd] |
Detailed help for a command |
Commands can be disabled in config if another service (like MeshMonitor) handles them.
Mesh Intelligence
MeshAI continuously polls mesh data sources and computes a 5-pillar health score:
| Pillar | Weight | What It Measures |
|---|---|---|
| Infrastructure | 30% | Router uptime — how many infra nodes are online |
| Utilization | 25% | Channel busyness — RF congestion across the mesh |
| Coverage | 20% | Gateway reach — how many monitoring sources see each node |
| Behavior | 15% | Traffic patterns — detecting noisy or misconfigured nodes |
| Power | 10% | Battery health — infrastructure nodes only |
Health Display
!health shows a compact overview with personality:
📡 Mesh 🟢 healthy
🏗️ 15/16 routers up
❌ Down: TVM Tablerock Relay
📶 152 full coverage, 94 on thin ice
🔥 Hayden Peak Router at 21% util
🔋 All infra powered ✅
🌡️ 29-34°C across 2 sensors
Treasure Valley 🟢 | Magic Valley 🟢
Status dots: 🔵 perfect (100) · 🟢 healthy (75+) · 🟠 warning (50+) · 🔴 critical (<50)
Monitoring Rules
Infrastructure nodes (routers, repeaters) are monitored individually with full detail — battery, offline alerts, coverage, neighbors, hardware. Client nodes dying is normal and not tracked. Channel utilization and environmental sensors are monitored for all nodes.
Conversational Queries
Ask questions naturally over LoRa:
- "how's the mesh?" → health overview with top issues
- "tell me about MHR" → full node detail with neighbors, coverage, feeders
- "where do we need more coverage?" → named gaps with specific nodes
- "how far is MHR from AIDA?" → GPS distance calculation
- "which nodes only reach one gateway?" → named nodes with their gateway
- "which gateway has the best signal?" → feeder comparison
Geographic Regions
Regions are configurable with local names, descriptions, aliases, and cities — all manageable through the TUI. No hardcoded geography in the code.
mesh_intelligence:
regions:
- name: "South Central ID"
local_name: "Magic Valley"
description: "Twin Falls area"
aliases: ["southern Idaho", "magic valley"]
cities: ["Twin Falls", "Burley", "Jerome"]
lat: 42.5
lon: -114.5
radius_km: 80
Data Sources
MeshAI aggregates from multiple sources using staggered tick-based polling (one API call per 30-second tick):
Meshview
Unauthenticated REST API. Supports multiple instances.
| Endpoint | Interval | Data |
|---|---|---|
/api/packets |
30s | Near real-time packet feed |
/api/nodes |
2 min | Node list with metadata |
/api/stats |
3 min | Traffic statistics |
/api/edges |
3 min | Node-to-node connections |
/api/traceroutes |
5 min | Route data |
/api/packets_seen |
10 min | Per-gateway RSSI/SNR (sampled) |
MeshMonitor
Authenticated (Bearer token). Single instance.
| Endpoint | Interval | Data |
|---|---|---|
/api/v1/packets |
60s | Packet feed |
/api/v1/nodes |
2 min | Nodes with battery, utilization, hardware |
/api/v1/telemetry |
2 min | Environmental sensors, device metrics |
/api/v1/traceroutes |
5 min | Route data |
/api/v1/channels |
5 min | Channel configuration |
/api/v1/network |
5 min | Network statistics |
/api/v1/solar |
10 min | Solar estimates |
Rate Limiting
Built-in protection for all sources: HTTP 429 backoff with Retry-After, exponential backoff on consecutive errors, slow response warnings, and optional polite mode for shared instances.
Source Configuration
mesh_sources:
- name: "local-meshview"
type: meshview
url: "http://192.168.1.100:8080"
enabled: true
- name: "meshmonitor"
type: meshmonitor
url: "http://192.168.1.100:3333"
api_token: "your-bearer-token"
enabled: true
Knowledge Base (RAG)
MeshAI uses a hybrid knowledge retrieval system with two backends:
Primary: RECON Qdrant Backend
Queries RECON's knowledge extraction pipeline — 2.8M+ vectors covering survival skills, communications, medical, technical documentation, Meshtastic docs, and more. Uses the same embedding infrastructure as RECON:
- Dense embeddings: TEI service with BAAI/bge-m3 (1024-dim)
- Sparse embeddings: bge-m3-sparse with IDF modifier
- Search: Qdrant hybrid with Reciprocal Rank Fusion (dense + sparse)
No data is copied — MeshAI queries RECON's Qdrant and TEI services over the network.
knowledge:
enabled: true
backend: auto # "qdrant", "sqlite", or "auto" (try qdrant, fall back)
qdrant_host: "192.168.1.150"
qdrant_port: 6333
qdrant_collection: "recon_knowledge_hybrid"
tei_host: "192.168.1.150"
tei_port: 8090
sparse_host: "192.168.1.150"
sparse_port: 8091
use_sparse: true
top_k: 5
Fallback: Local SQLite
If the Qdrant backend is unreachable, MeshAI falls back to a local SQLite knowledge base using FTS5 keyword search and bge-small-en-v1.5 vector embeddings (384-dim).
# Build from Meshtastic ZIM file
python scripts/zim_to_knowledge.py meshtastic.zim --output knowledge.db
knowledge:
enabled: true
backend: sqlite
db_path: /data/meshai_knowledge.db
top_k: 5
Requires sqlite-vec and fastembed for the SQLite backend.
Architecture
┌──────────────────────────────────────────────────────────────────────┐
│ MeshAI │
├──────────────────────────────────────────────────────────────────────┤
│ │
│ DATA SOURCES INTELLIGENCE DELIVERY │
│ ┌─────────────┐ ┌──────────────┐ ┌────────────┐ │
│ │ Meshview ×N │─────┐ │ Health Engine │────────▶│ Reporter │ │
│ │ (staggered) │ │ │ 5-pillar │ │ Tier 1/2 │ │
│ └─────────────┘ ▼ │ scoring │ └─────┬──────┘ │
│ ┌─────────────┐ ┌──────┴──┐ │ │ │
│ │ MeshMonitor │─▶│ Data │─┘ │ ┌─────▼──────┐ │
│ │ (staggered) │ │ Store │ │ │ Router │ │
│ └─────────────┘ │ SQLite │ │ │ scope/dist │ │
│ └─────────┘ │ └─────┬──────┘ │
│ │ │ │ │
│ ┌────▼────┐ ┌─────▼──────┐ ┌────▼────┐ │
│ │ Feeder │ │ LLM │ │ Chunker │ │
│ │ Sampling│ │ Backend │ │LoRa-fit │ │
│ └─────────┘ └────────────┘ └────┬────┘ │
│ │ │
│ KNOWLEDGE ALERTS DELIVERY │ │
│ ┌─────────────┐ ┌─────────────┐ ┌──────────────┐ │ │
│ │ RECON/Qdrant│ │ Alert │ │ Subscription │ │ │
│ │ 2.8M vectors│ │ Engine │ │ Manager │ │ │
│ │ (network) │ │ 17 triggers │ │ daily/weekly │ │ │
│ ├─────────────┤ │ scaling │ │ alerts │ │ │
│ │ SQLite FTS5 │ │ cooldown │ └──────┬───────┘ │ │
│ │ (fallback) │ └──────┬──────┘ │ │ │
│ └─────────────┘ │ ┌─────▼────────┐ │ │
│ └─────────▶│ Responder │◀──┘ │
│ ┌─────────────┐ │ ACK-paced DM │ │
│ │ Conversation│ │ Channel alert│ │
│ │ History │ └──────────────┘ │
│ └─────────────┘ │
│ │
└───────────────────────────────────────────────────────────────────┘
│ │
┌────▼────┐ ┌────▼────┐
│ TEI │ │ Qdrant │
│ bge-m3 │ │ hybrid │
│ cortex │ │ cortex │
└─────────┘ └─────────┘
Message Chunking
Long responses are split into mesh-friendly chunks with sentence-aware splitting, configurable limits, and continuation prompts. Command output (like !health) packs multiple lines into 2-3 messages using newlines within each message to minimize airtime usage.
response:
max_length: 200 # Max chars per message
max_messages: 3 # Messages before continuation prompt
Alerting
Real-time alerts when mesh conditions change, with scaling cooldowns to prevent spam.
Alert Conditions (17 total, each toggleable)
| Pillar | Condition | Default Threshold |
|---|---|---|
| Infrastructure | Router goes offline | — |
| Infrastructure | Router recovery | — |
| Infrastructure | New router appears | — |
| Power | Battery warning | <50% |
| Power | Battery critical | <25% |
| Power | Battery emergency | <10% |
| Power | 7-day declining trend | >10% drop with rate |
| Power | USB → battery (power outage) | — |
| Power | Solar not charging during day | — |
| Utilization | Sustained high utilization | >20% for 6h |
| Utilization | Packet flood | >500 pkts/24h |
| Coverage | Infra drops to single gateway | — |
| Coverage | Feeder gateway stops responding | — |
| Coverage | Region total blackout | All infra offline |
| Scores | Mesh health score drop | <70/100 |
| Scores | Region health score drop | <60/100 |
Scaling Cooldown
Alerts don't spam. When a condition triggers:
- Alert 1: fires immediately
- Alert 2: 12 hours later (if still in condition)
- Alert 3: 24 hours after that
- Alert 4: 48 hours after that
- Stops until condition resolves
When the condition clears, one recovery notification fires and the tracker resets.
Delivery
Alerts are delivered two ways:
- Channel broadcast: configurable channel index for mesh-wide visibility
- DM to subscribers: users who ran
!sub alertsreceive DMs matching their scope
Critical Nodes
Designate important infrastructure (e.g., MHR, HPR) as critical. When a critical node goes offline, alerts use priority formatting.
mesh_intelligence:
critical_nodes: ["MHR", "HPR"]
alert_channel: 0 # Channel for broadcast alerts (-1 = disabled)
All conditions and thresholds are configurable via the TUI under Mesh Intelligence → Alert Rules.
LLM Configuration
llm:
backend: "google" # openai, anthropic, google
api_key: "your-api-key"
model: "gemini-2.0-flash"
Local LLMs
MeshAI works with any OpenAI-compatible API:
- LiteLLM:
base_url: "http://localhost:4000/v1" - Open WebUI:
base_url: "http://localhost:3000/api" - Ollama:
base_url: "http://localhost:11434/v1"
Docker
TCP Connection (recommended)
connection:
type: "tcp"
tcp_host: "192.168.1.100"
tcp_port: 4403
Serial Connection
connection:
type: "serial"
serial_port: "/dev/ttyUSB0"
Edit docker-compose.serial.yml to match your device path.
Environment Variables
LLM_API_KEY=your-key-here docker compose up -d
Running Alongside Other Services
advBBS
MeshAI coexists with advBBS on the same node. BBS protocol messages (sync, RAP, mail notifications) are automatically filtered. No configuration needed.
bot:
filter_bbs_protocols: true
MeshMonitor
MeshAI integrates with MeshMonitor at two levels: it fetches MeshMonitor's auto-responder patterns to avoid duplicate responses, and it uses MeshMonitor's API as a data source for mesh intelligence (battery, telemetry, traceroutes, solar).
meshmonitor:
enabled: true
url: "http://192.168.1.100:8080"
inject_into_prompt: true
refresh_interval: 300
Running as a Service
# /etc/systemd/system/meshai.service
[Unit]
Description=MeshAI - Meshtastic Mesh Intelligence
After=network.target
[Service]
Type=simple
User=your-user
WorkingDirectory=/path/to/meshai
ExecStart=/usr/bin/python3 -m meshai
Restart=always
RestartSec=10
[Install]
WantedBy=multi-user.target
sudo systemctl daemon-reload
sudo systemctl enable meshai
sudo systemctl start meshai
Acknowledgments
- Meshtastic — the mesh networking platform
- MeshMonitor by Yeraze — monitoring integration and data source
- advBBS — BBS coexistence design
- sqlite-vec by Alex Garcia — vector search in SQLite
- fastembed by Qdrant — fast local embeddings
License
MIT License
Author
K7ZVX - matt@echo6.co