You're running SeekDeep through the desktop shell — the .msi already did steps 1-8 for you. The local AI server is up, Python deps are installed, and the tray icon manages everything. Every "Install" button below is misleading for your install (you already have Node, Python, etc — that's how this page is even loading). Free to keep scrolling out of curiosity; nothing here is required.
SeekDeep runs entirely on your own machine. This wizard will check your system, install the bot, configure Discord, download models, and bring up the full local stack. Nothing leaves the box.
REQUIRED
Node 20+
Discord bot runtime · package.json engines
REQUIRED
Python 3.10+
Local AI server (FastAPI)
REQUIRED
NVIDIA 12GB+
VRAM for chat + image
RECOMMENDED
Docker + compose
SearXNG container
RECOMMENDED
~ 80 GB
SSD for model cache
REQUIRED
Discord app + bot
token + client ID
▸ ETA
About 15-20 minutes including downloads. Most of that is the first model pull (~80 GB across 6 models). You can re-run this wizard later to add models or change configuration.
▸ "Windows protected your PC" warning · what it means
If Windows showed you the blue "Microsoft Defender SmartScreen prevented an unrecognized app from starting" warning when you launched SeekDeep, that's expected. The build is unsigned — getting a Microsoft-recognized code-signing certificate costs $300–$700 a year, and even then SmartScreen needs to see a few hundred installs before it stops nagging. SeekDeep is a one-developer open-source project; paying that toll isn't viable.
To run it past SmartScreen: click More info → then the Run anyway button that appears. SmartScreen remembers your choice — you won't be asked again for the same installer.
Drop the .msi into VirusTotal for a 70-engine scan — typically clean across the board.
Network calls are documented at the top of the README: nothing phones home; only huggingface.co for model downloads (opt-in), 127.0.0.1 for your own SearXNG/Ollama, and any remote chat endpoint you explicitly configure.
Verify your environment can run SeekDeep. Click Run all checks below — SeekDeep will probe Node, Python, Docker, GPU, disk space, and any already-running services.
Node.jsruntime for index.js · 20+ required (package.json engines)
NVIDIA GPU + CUDAprobed via live AI server · install + warm models if missing
—
via /gpu
Disk space~ 80 GB for model cache + venv + node_modules
—
manual
Local AI serverlive probe: GET http://127.0.0.1:7865/health
—
probe
SearXNGlive probe: GET http://127.0.0.1:8080
—
▸ NOTE
Browser-only checks reach localhost services via fetch(). Node / Python / Docker / disk readings come from the AI server's /health when it's running. If it isn't running yet, that's fine — they go yellow ("install + come back"). The first run is always cold.
— equivalent to npm run doctor · streams live
Sets up everything setup_local.ps1 + npm install used to require a terminal for: Python .venv, requirements-local.txt, working directories, a starter .env, and node_modules. One click. No PowerShell.
PROBING …
— skips anything already in place
Bootstrap complete
Auto-toggles green when /system/bootstrap-status reports every row ✓. You don't have to flip this manually.
Paste your bot's Discord token below. SeekDeep stores it locally — only your machine ever sees it. If you don't have one yet, follow the dev portal steps in the hint.
▸ Create one at discord.com/developers/applications → New Application → Bot → Reset Token. Required scopes: bot + applications.commands. Permissions integer: 8590094848. Token shape: starts with M or N + 23ish chars, two dots, base64-ish blocks.
▸ Same app page → General Information → copy the Application ID. Required for slash-command registration. npm run doctor will fail on a placeholder value.
▸ Comma-separated Discord user IDs. Admins get priority on the image queue + persona / digest perms. Right-click yourself in Discord → Copy User ID (Developer Mode must be on).
▸ Only needed if you're using gated models like Llama-3.1-8B-Instruct. Get one at huggingface.co/settings/tokens and accept each model's license on its page first. Validated client-side — must start with hf_.
▸ Pick one (or both). LOCAL is a daemon at 127.0.0.1:11434, no auth. CLOUD uses your ollama.com account — either paste an API key from Account → Keys, OR click "Sign in" to run ollama signin which uploads a device ed25519 key (the "Device Keys" section in the portal).
— not yet probed
Pick the models SeekDeep will load. Defaults are tuned for a 24GB GPU. Each model loads only when the matching task fires; the task-LRU keeps just one chat model resident.
▸ Cards below are pulled from the live catalog (/models/catalog) on page load — HF + Ollama, sized for your detected VRAM. If the AI server is unreachable, you'll see a curated fallback set. More install actions + on-disk model presence live in Control Center → Model manager.
▸ DEFAULT_CHAT · pick one
8B · 4bit5.1 GB
Llama-3.1-8B-Instruct
meta-llama/Llama-3.1-8B-Instruct
12B · 4bit7.2 GB
Mistral-Nemo
mistralai/Mistral-Nemo-Instruct-2407
8B · 4bit4.9 GB
Granite 3.3
ibm-granite/granite-3.3-8b-instruct
▸ VISION · pick one
3B · fp163.4 GB
Qwen2.5-VL-3B
Qwen/Qwen2.5-VL-3B-Instruct
7B · fp167.8 GB
Qwen2.5-VL-7B
Qwen/Qwen2.5-VL-7B-Instruct
SKIP0 GB
No vision
disables /vision + OCR mode
▸ IMAGE GEN · pick one
SDXL · fp166.8 GB
Dreamshaper-XL
Lykon/dreamshaper-xl-1-0
SDXL · fp166.6 GB
SDXL Base 1.0
stabilityai/stable-diffusion-xl-base-1.0
SKIP0 GB
No image gen
disables /image + edit pipelines
VRAM BUDGET · TASK-LRU SWAP14.3/ 24 GB
chat 5.1image 6.8vision 3.4os reserve 4.0
LOCAL_CHAT_QUANT = 4bit
4-bit NF4 quantization · ~1-2% quality drop, fits 12-14B models on 24GB. Recommended on consumer GPUs.
SearXNG is a privacy-respecting metasearch engine. SeekDeep uses it for web:auto and web:always routes. It runs in a single docker container on 127.0.0.1:8080.
— starts the docker container, no terminal
SearXNG live statusfetch http://127.0.0.1:8080
not yet probed
WEB_AUTO_SEARCH
Let SeekDeep auto-decide when to call SearXNG for fresh-data questions.
WEB_SEARCH_FAIL_OPEN
If SearXNG is down, fall back to model knowledge instead of erroring.
Pre-download the models you selected. This is the slowest step — about ~ 15 GB for the defaults. After the first warmup, SeekDeep boots from the local cache without internet.
— pulls weights into ./models/huggingface · streams live
▸ WARM downloads everything from warmup_local_cache.py — chat, image, vision models per .env. Slowest step (10-20 min on 100 Mbps). ▸ LOCK flips HF_HUB_OFFLINE=1 in .env so subsequent boots refuse to hit the hub. Tap Lock only AFTER warm completes.
Auto-toggles when warmup_local_cache.py exits with code 0.
Bring the stack up and run the smoke test. seekdeep_launcher.bat option 8 starts SearXNG, the Python AI server, and the Node bot in dependency order.
▸ LAUNCH FULL STACK sequences SearXNG → AI server → Discord bot just like seekdeep_launcher.bat option 8 used to. ↻ PROBE hits each check row below to verify reachability. All without a terminal.
Local AI server · /healthGET http://127.0.0.1:7865/health
—
GPU snapshot · /gpuGET http://127.0.0.1:7865/gpu
—
SearXNG · web searchGET http://127.0.0.1:8080
—
Discord bot · gatewayverified once you see SeekDeep online in your guild
—
SeekDeep is online in my Discord server
Last manual confirmation. Once you see the bot online, flip this on.
Setup complete.
Your configuration is ready. Two ways to apply it:
APPLY TO .env — writes it directly via the local AI server. Fastest, no copy-paste.
SHOW PREVIEW — see the raw .env block and copy-paste it yourself.
Either way, the wizard is re-runnable any time. Hit Finished at the bottom to open the Control Center.