A powerful, stateful, and lightning-fast personal assistant for Discord. Powered by local Ollama (qwen3:8b) and built with a modular, tool-intelligent architecture.
Caution
Self-Bot Warning: Automating user accounts is strictly against Discord's Terms of Service. Using this tool carries a risk of permanent account suspension. Use at your own discretion.
The assistant uses a self-directed ReAct loop — on each message it decides its own next step:
- Respond directly for simple conversational queries.
- Engage Thinking Mode (
[MODE: think]) for deep reasoning, debugging, or complex analysis. - Call a tool (
[ACTION: tool(args)]) when it needs external data, and loop back with the result.
No separate routing model needed — intent is determined dynamically by the model itself.
- ⛅ Live Weather: Instant reports for any city via
wttr.in. - 🔢 Safe Calculator: Accurate math without LLM hallucination.
- ⏰ Smart Reminders: Precise timers and alerts.
- 🔍 Deep Search: Real-time web search via your local SearXNG instance.
- 🧠 Persistent Memory: Private database that remembers facts about you across sessions.
- 📊 Stats Dashboard: Token usage, searches, and tool executions.
Attach and analyse plain-text files directly in Discord. Supported formats:
.txt .log .md .rst .csv .tsv .json .yaml .yml .toml .ini .cfg .conf .env .xml .html .htm .css .diff .patch .ttml
Note
Executables, scripts, binaries, and media files are blocked for security. Up to 5 files × 200 KB per message.
Send images and get intelligent analysis powered by a vision model or OCR fallback.
Option A — Vision Model (recommended): Uses a local Ollama vision model (e.g. moondream, llava, qwen2-vl) to fully describe images, transcribe text, and interpret screenshots.
Option B — pytesseract OCR: Text extraction only, no extra model needed. Good for screenshots with dense text.
Supports: .png .jpg .jpeg .webp .gif — up to 3 images × 10 MB per message.
- Current day, date, and UTC time injected into every response.
- Channel history summarisation for long conversations.
- User profile awareness (bio, pronouns, roles, rich presence / RPC).
Custom Discord Rich Presence with the official Google Gemini icon, showing live stats (messages answered, tokens used) while keeping the assistant running on your account.
ollama pull qwen3:8b
# Optional: vision model for image understanding
ollama pull moondreampip install -r requirements.txt
# Optional: OCR fallback (if not using a vision model)
pip install pytesseract pillow
sudo pacman -S tesseract tesseract-data-eng # Arch / CachyOS
# sudo apt install tesseract-ocr # Debian / Ubuntu# --- Required ---
DISCORD_TOKEN=your_token_here
OLLAMA_MODEL=qwen3:8b
SEARXNG_BASE_URL=http://localhost:8888
# --- Context Window ---
# Larger = more memory, better recall. Tune to your VRAM.
# Approx VRAM cost on qwen3:8b (Q4): 8k≈7.8GB, 16k≈8.9GB, 32k≈11GB
OLLAMA_NUM_CTX=32768
# --- Image Support (optional) ---
# Set to a vision-capable model name, or leave empty to use pytesseract OCR fallback
OLLAMA_VISION_MODEL=moondream
# GPU layers for the vision model:
# 0 = fully on CPU/RAM — recommended, keeps VRAM free for main model
# -1 = fully on GPU — fastest, but evicts main model from VRAM
VISION_NUM_GPU=0python main.pyPrefix or suffix your message to force a specific path:
| Flag | Behaviour |
|---|---|
--think |
Activates extended reasoning mode |
--search |
Forces a web search |
--stats |
Shows the performance dashboard |
OLLAMA_NUM_CTX |
VRAM (qwen3:8b Q4) |
|---|---|
| 8 192 | ~7.8 GB |
| 16 384 | ~8.9 GB |
| 32 768 | ~11.0 GB |
Vision model co-existence: With VISION_NUM_GPU=0, the vision model runs entirely in system RAM — no VRAM is consumed and the main model is never evicted. Any system with 8+ GB RAM can run moondream this way alongside a 32k-context qwen3:8b.
This project is licensed under the MIT License — see the LICENSE file for details.