SeekDeep · v— · local-first Discord assistant

SeekDeep

A Discord bot that runs Llama-3.1, Qwen-VL, and SDXL on your own GPU. Includes archive threads, persona overrides, role-routed chat, image-edit pipelines, and live VRAM monitoring — with no outbound API calls.

Open Control Center Installer
Offline capable Windows-first No telemetry GPL-2.0
N · 000 S · 180 W · 270 E · 090 SeekDeep mark

Design principles

01 / why local
01 · Private by default

Prompts stay on the host.

No OpenAI, no Anthropic, no telemetry. Chat history, archive threads, generated images, and search queries remain on the machine running the bot.

0
Outbound API calls
02 · Routed by task

One model per role.

Routes pin specific models: quality chats run Mistral-Nemo, reasoning runs Phi-4, lightweight fallback runs Gemma-3n. A task-LRU policy bounds VRAM use.

5
Chat-role models
03 · VRAM-aware

Bounded by the GPU budget.

Live GPU monitor, thrashing warnings, 4-bit quantization, configurable system reserve, and a single-flight lock for safe model swaps under load.

24 GB
Reference target
Chat

Mentions, slash commands, and web-routed answers.

Trigger with @SeekDeep or a slash command. web:auto defers to the router; web:always forces a SearXNG round-trip. Conversation context is held in a rolling buffer, scoped per user and per channel.

  • @SeekDeep ask, refine, translate, persona
  • /help · /status · /search · /recent · /stats
  • Auto-translate channels for non-Latin scripts
  • Daily digest channel · changelog · ping
N
nautilus
@SeekDeep what's the latest on the Kraken expedition findings?
SeekDeep · WEB ROUTED · 1.2s
The 2026 NOAA survey logged 47 new colossal-class specimens at the Tonga Trench between 8,400m and 10,200m. Sources: <noaa.gov/...> <nature.com/...>
N
nautilus
@SeekDeep show me a colossal squid at 10000m, raw
SeekDeep · QUEUE 1/1 · STEPS 28
▸ Original · Refined · Both?
Image pipeline

SDXL with chained edits.

Dreamshaper-XL at 28 steps with the dpmsolver++ scheduler and prompt refinement via the pinned chat model. Edits chain through img2img, InstructPix2Pix, CLIPSeg-masked inpaint, and Lanczos upscale.

  • @SeekDeep show me · draw · render · paint
  • img2img · pix2pix · inpaint · upscale 2x/3x/4x
  • 10 style presets · 3 quality tiers
  • Archive button · shared archive · re-refine
SEED 8842 · STEPS 40 · 1024²
img2img · 0.62
inpaint · clipseg
VRAM monitor

Live GPU accounting.

Every command reports GPU and VRAM use. When chat, image, and vision can't co-reside, the task-LRU evicts the coldest model and surfaces the decision in the response.

  • @SeekDeep gpu watch [N] · live tail
  • System reserve · safety margin · singleflight
  • 4-bit chat quant by default · fp16 opt-in
  • Pin chat or vision · explicit /unload
RTX 4090 · LAPTOP · 24GB
14.2/ 24.0 GB
default_chat
5.1 GB
image / sdxl
6.8 GB
vision
3.4 GB
system reserve
4.0 GB
HEALTHY last evict: 4m ago · vision
Setup · Windows 11 · ~6 minutes

Four steps.

01

Clone

Drop the repo wherever you want it. No system installs.

git clone seekdeep && cd seekdeep
02

Setup

PowerShell does the venv, the npm install, and copies .env.default to .env for you.

./setup_local.ps1
03

Token

Paste your bot's DISCORD_TOKEN into .env. Optional admin IDs go below it.

DISCORD_TOKEN=...
04

Launch

The launcher brings up SearXNG, the AI server, and the bot in the right order.

./seekdeep_launcher.bat
SeekDeep · v— · local Open source · GPL-2.0 licensed