Epicurious — The Open AI Agent Studio & Skill Forge

⌘K

Theme

Accent color

Agents

Skills

279

Commands

Active Model

llama3.2:3b

Name	Source	Size	Files	Verdict	Imported	Actions

Drop .md/.txt files to import, or click to browse (multiple supported)

		Name	Lang	Size	Date	Actions

Settings

AI Models

LLM Models

Embedding Models

Model Wizard

Integrations

MCP Servers

SearXNG Engine

Unlimited-OCR

Audio Transcription

System

Monitoring

Install

Preferences

Appearance

Select Active LLM Generation Model

Choose the primary model used by Epicurious subagents and chatbot generation.

Add More Models

Scan an endpoint (Ollama, LM Studio, vLLM, etc.) and review/select what to add

llama3.2:3b

Local

Local Ollama model (localhost:11434) — used for chat, SQL generation, and wizard content.

Selecting a model applies and saves it immediately — no separate save step required.

Looking for custom prompt instructions? Open the chat panel's own Settings (Personalization > Custom Prompts).

Unlimited-OCR Vision Control

Enable or disable background OCR / vision model processing to manage GPU VRAM allocation.

Service & VRAM Management

Endpoint: http://192.168.2.10:8200
Model: baidu/Unlimited-OCR (vLLM Engine)
Behavior: When enabled, scanned PDFs and image files uploaded via Document-to-Skill or Chat will route through Unlimited-OCR for layout text extraction. Keep OFF when not needed to save VRAM for LLM models.

Unlimited-OCR Vision Fallback Status

Currently OFF — vision OCR will not run on uploaded PDFs/images

SearXNG Web Search Engine

Configure and monitor live web search integrations via your SearXNG instance.

Connection Details

Container: onyx-searxng on host spark:8888
Config Path: spark:/home/gigagrid/searxng/config/settings.yml
API Route: /api/searxng/search?q=...

Enabled Search Engines

✓ DuckDuckGo ✓ Google ✓ Brave ✓ Bing ✓ Wikipedia

Select Active Embedding Model

Choose vector embedding models for codebase semantic search and RAG indexing.

Nomic Embed Text

Local

8192 context window local embedding model for code & docs.

Text Embedding 3

OpenAI

High-density 3072 dimension vector embedding API.

BGE Large v1.5

Local

BAAI top-ranking retrieval & semantic similarity model.

Selecting a model applies and saves it immediately — no separate save step required.

Model Context Protocol (MCP) Server Connections

Connect to local stdio, SQLite databases (WhatsApp context DB, CRM), PostgreSQL, or remote SSE/HTTP Model Context Protocol servers.

Quick Source Presets

Active Connected Custom MCP Servers

Audio Transcription Engines & Models

Manage local audio transcription engines, download Whisper & Moonshine models, and configure default transcription servers.

Active Local Engine Endpoint

Server: http://127.0.0.1:8103 (whisper.cpp & Moonshine engine)
Right-Click Quick Transcribe: Right-click any .mp3, .wav, .m4a, .mp4, .aac, .ogg, .oga, .opus, .flac, .webm, .mov, or .wma file in the file explorer to transcribe immediately. Also available from the chat panel's mic button (🎙️).

Available Local Whisper Models (whisper.cpp)

Moonshine On-Device Models

Add Models From an Endpoint

Point at an IP/host running Ollama or an OpenAI-compatible server (LM Studio, vLLM, etc.) — it actually queries the server, lists every real model found, and you pick which ones to add.

Endpoint (IP:port or URL)

Appearance

Choose a color theme for the dashboard.

Default

Warm Orange

Dark, warm-orange accent theme.

Bombay Sapphire

Liquid Glass

Deep sapphire blue with a glowing cyan accent and animated liquid-glass surfaces.

Evergreen

Jewel Beetle

Near-black with a metallic forest-green sheen and bright gold spot-color accents.

First-Run Setup

Backend Monitoring

Live status of every service the dashboard talks to. Auto-refreshes every 15s.

Last checked

Never

Install Info

System info & quick commands.

Install location

~/.claude/

Source repo

~/Epicurious/

⚡ FastMCP CLI Installation

Use the fastmcp CLI tool to automatically install and register this harness into your favorite AI client with one command:

Claude Desktop

fastmcp install claude-desktop /Users/guychenya/Epicurious/scripts/epi.js

Cursor IDE

fastmcp install cursor /Users/guychenya/Epicurious/scripts/epi.js

Claude Code CLI

fastmcp install claude-code /Users/guychenya/Epicurious/scripts/epi.js

Manual JSON setup (for claude_desktop_config.json or antigravity.json):

{
  "mcpServers": {
    "epicurious-harness": {
      "command": "node",
      "args": ["/Users/guychenya/Epicurious/scripts/epi.js", "mcp-server"]
    }
  }
}

Quick commands

node ~/Epicurious/scripts/epi.js doctor

node ~/Epicurious/scripts/epi.js list-installed

Update Epicurious

cd ~/Epicurious && git pull && bash ./install.sh --profile full --target claude

Chat Spaces

Pick a folder for Chat's Docs mode to search directly. If the folder contains a ZIP archive (PDF, DOCX, PPTX, XLSX, EPUB, or TXT documents inside), an Unzip action appears in that space's row to extract and convert it to Markdown.

Add a Folder Space

Chat's Docs mode will search this folder directly — no upload needed.

Configured Spaces

Analytics

Overview

Main KPIs

Activity

Recent Activity

Agent Chain Traces

Chat File Access

Loading analytics...

Main KPIs

Capability composition, wizard activity, and the SkillSpector security gate at a glance. Color signals state: green is healthy, amber needs review, red is a problem.

Live · updated just now

Total Capabilities

—

Security Gate Health

Compositioncurated vs. imported

Wizard Activitylifecycle events

Trend deltas are illustrative placeholders showing where week-over-week change would appear once history is wired in.

SkillSpector Security Gate

Top Contributors

System Health

Ask a Question (SQL Chat)

Local LLM writes a read-only query against the wizard database (imported_skills, activity_log, curated_certifications). No writes are ever permitted.

Recent Activity

Every import, delete, disable, block, and direct file edit the wizard has recorded.

When	Action	Kind	Name	Author	Verdict

Agent Chain Traces

Every chat-driven @agent/@command run — tag one for a single execution or two+ for a sequential chain. Click a row to see each step's prompt, tool calls, and output.

When	Question	Chain	Status

Chat File Access

Every file chat has read to answer a question, with the question asked and when. File contents and AI answers are never logged here — path and timestamp only.

When	File	Question