undesirables-mcp-server

<!-- mcp-name: io.github.sailorpepe/undesirables-mcp-server --> <div align="center">

35+ local compute tools for AI agents — plus a hosted, conformal-calibrated pricing oracle

Run it fully local with zero cloud dependency, or connect to the hosted endpoint at mcp.the-undesirables.com — pay-per-call in USDC, no account or API key.

Website · Docs · PyPI · 𝕏

</div> <div align="center"> <img src="https://raw.githubusercontent.com/sailorpepe/undesirables-mcp-server/main/assets/demo.gif" alt="Undesirables MCP Server Demo" width="480" /> </div>

---

🔌 Connect over MCP — one URL, no install

https://mcp.the-undesirables.com

Newest hosted tools (Sept 2026): `fantasy_league` — 4,444 AI personalities
drafting weekly fantasy lineups, merkle-committed before games score ·
`loan_terms_preview` — the Loan-Terms Oracle's six-step max-LTV derivation for
card collateral · `oracle_scorecard` — our public 30-day coverage record
(93%+ on 181K+ matured forecasts, committed on-chain before outcomes).

No install, no account, no API key. 23 tools over streamable HTTP (MCP protocol 2025-06-18; legacy SSE also served). Free tools answer immediately. Paid tools return an x402 payment_required carrying amount, network, and payTo — an agent with a funded wallet can settle and retry in the same session. Settlement only occurs on a successful response; failed calls are never charged.

Holders can also ask which Undesirables a wallet owns and how each soul's calls have scored (souls_in_wallet, soul_calls) — public track record only; personalities stay holder-gated.

Claude Desktop / Perplexity — add it as a custom remote connector (Perplexity: Settings → Connectors → + Custom Connector → Remote).

Cursor / Windsurf / VS Code — clients that take a URL in config:

{
  "mcpServers": {
    "tcg-oracle": { "url": "https://mcp.the-undesirables.com" }
  }
}

Tools: search_tcg_products, market_snapshot, grade_card, grade_or_not, simulate_price, card_forecast, trending_cards, optimize_portfolio, recommend_workflow, check_accuracy.

Search is set-aware — search_tcg_products("Base Set Charizard") separates Base Set, Base Set 2, and Shadowless rather than returning every Charizard printing. Every result carries a set field and a product_id you can pass straight to the other tools.

---

Quick Install

pip install undesirables-mcp-server

Turn any Undesirable NFT into an MCP-compatible AI agent with 35+ local compute tools.

---

Table of Contents

• What It Does
• What's New in v1.1.8
• Prerequisites
• Full Setup
• Boot The Server
• Connect Your Chat Front-End
• Local Image Generation
• Troubleshooting
• Technical Architecture
• Agent Framework Integration
• LitVM TCG Oracle
• Ecosystem
• License & Commercial Use

---

What It Does

• 🎴 Vision AI Card Grading — PSA/Beckett prediction via Qwen VL
• 📊 Conformal Risk Forecast — calibrated VaR/CVaR + Safe-Hold & Momentum letter grades (Monte Carlo GBM/Merton opt-in)
• 🎵 AI Music Generation — ACE Step on Apple Silicon
• 🎬 Video Clipping & Beat Sync Editing — FFmpeg
• 🖼️ Local Image Generation — MLX Flux on Mac, DirectML on Windows, CUDA on Linux
• 🗣️ Text to Speech Voice Engine — Kokoro TTS
• 🧠 Persistent RAG Memory Graphs — CRM node mapping
• 🔍 Zero Token Web Search — DuckDuckGo
• 🔒 SAST Code Security Auditing
• 📈 Financial Analytics Oracle — TCGCSV + eBay depth analysis

---

<details> <summary><strong>What's New in v1.1.8</strong></summary>

v1.1.8 adds the FREE card_forecast(card_name | product_id) tool — one call returns the conformal 30-day price forecast plus Safe-Hold & Momentum letter grades and a one-line plain-English read (e.g. "~12% chance it's below $Y in 30 days; Safe-Hold B, Momentum A"). No payment required.

The conformal-calibrated risk forecast is the default model — regime-aware split-conformal bands with honest VaR/CVaR, plus Safe-Hold & Momentum letter grades. Monte Carlo (GBM / Merton Jump-Diffusion) remains available opt-in via model=. Also: corrected license badge and full ecosystem integration.

Key Features:

• purchase_undesirables_license_key — Returns an unsigned EVM transaction payload (Ethereum Mainnet, chainId 1) for autonomous agents to mint directly from the Scatter.art contract
• verify_soul_initialization — Verifies on chain purchase via public RPC and initializes the cryptographic soul matrix, unlocking all local compute engines
• Verified on Glama.ai with a 3.8/5 quality score across 36 tools
• Listed on 9+ MCP directories including the Official MCP Registry

</details>

---

🛑 Prerequisites (Read Carefully)

If you've never used Python or run AI Models locally, you must do this first:

1. Download Python (Version 3.10 or higher).
2. Download Ollama. CRITICAL: You cannot just download the app and leave it in your downloads folder. You must double-click the Ollama app to physically run it. You should see a little llama icon in your Mac menu bar or Windows system tray for this server to work.

---

🛠️ Step 1: Install & Clone

First, open your Terminal or Command Prompt and clone this repository. After cloning, you must activate a "Virtual Environment" (a sandbox folder just for this codebase).

🍎 On Mac / Linux

git clone https://github.com/sailorpepe/undesirables-mcp-server.git
cd undesirables-mcp-server
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

🪟 On Windows

git clone https://github.com/sailorpepe/undesirables-mcp-server.git
cd undesirables-mcp-server
python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

---

🚀 Step 2: Boot The Server

Every single time you want to run this server later, you must open your terminal and make sure your Virtual Environment is activated (venv) first!

If you already downloaded your Soul Workspace from the website:

# Make sure to point to your EXACT soul folder path
python server.py --workspace "/Users/username/Desktop/soul_folder/0420"

---

🔌 Step 3: Connect Your Chat Front-End

The MCP Server doesn't have a chat window; it runs invisibly in the background of your terminal! To actually talk to your agent, you must connect it to a desktop application like Claude or Cursor.

Claude Desktop Connection

1. Open the Claude Desktop application on your computer.
2. Go to Settings > Developer > Edit Config.
3. Paste this into your config file, making absolutely sure you replace the cwd (Current Working Directory) with your exact folder path:

{
  "mcpServers": {
    "undesirables": {
      "command": "python",
      "args": ["server.py", "--workspace", "/Users/yourname/Desktop/soul_folder/0420"],
      "cwd": "/Users/yourname/Documents/undesirables-mcp-server"
    }
  }
}

4. Restart the Claude Desktop app. You should see a little "Plugin/Hammer" icon telling you that 35+ The Undesirables tools are now available!

---

🎨 Step 4: Setup Local Image Generation (Optional)

If you want your agent to physically generate memes and illustrations 100% offline natively on your computer, the MCP Server uses the massively powerful 16GB FLUX.1-schnell model.

If you do not complete this step, or if your computer is too weak (< 12GB RAM), the server will automatically fallback and generate memes for you silently via the free Pollinations.ai cloud network.

🍏 Authenticating Apple Silicon (Mac M1/M2/M3/M4)

Apple Silicon specifically uses mflux, which strictly requires a Hugging Face token to bypass Black Forest Labs' legal compliance gate.

1. Navigate to black-forest-labs/FLUX.1-schnell, create a free Hugging Face account, and click Agree and Access.
2. Go to Hugging Face Tokens and generate a new Read token.
3. Open your Mac terminal, activate your virtual environment, and log in:

cd undesirables-mcp-server
source venv/bin/activate
python -c "import huggingface_hub; huggingface_hub.login()"

4. Paste your token and press Enter (your clipboard characters will be invisible for security).

🪟 Setup for Windows/Linux GPUs

If your computer uses Nvidia CUDA or AMD DirectML, the diagnostic scanner detects this and logically shifts your engine to an ungated open-weights repository (shuttleai/FLUX.1-schnell).

• You do not need to authenticate anything or make an account.
• Simply ask your agent to generate a meme in the UI! Your system will natively download the 16GB weights fully offline during the very first execution automatically.

---

⚠️ Common Idiot-Proof Diagnostics

If your terminal throws red text and halts, check these top 3 reasons:

• Error: Ollama connection refused Your AI's brain is offline! Make sure you physically double-clicked the Ollama.app on your computer. If the little llama icon isn't in your menu bar/taskbar, local inference will fail immediately.

• ModuleNotFoundError: no module named fastmcp You forgot to activate your Virtual Environment. You cannot just launch a fresh terminal and run python server.py. You must navigate to the folder and run source venv/bin/activate (Mac) or venv\Scripts\activate (Windows) first!

• Invalid JSON: expected value at line 1 The Python terminal running the MCP Server is communicating in raw machine code (JSON-RPC). You cannot type plain English into that terminal window! Once it turns on, leave it alone. Open Claude Desktop or Cursor to chat with it.

---

Technical Architecture (For Developers)

This MCP server exposes your local NFT soul via the Model Context Protocol standard.

Resources (read only context your AI can access):

• soul://personality — Big Five scores, archetype, strategy, fatal flaw
• soul://system-prompt — The full system prompt that defines the agent
• soul://memory — Persistent memory (trade history, observations)
• soul://predictions — Prediction ledger with grades

Core Tools (35+ functions your AI can call):

• grade_tcg_card — 3-stage PSA/Beckett grading: Qwen Vision LLM + OpenCV centering + BGS capping
• card_forecast — FREE one-call conformal 30-day forecast + Safe-Hold/Momentum letter grades + plain-English read (pass a card name or TCGplayer product_id)
• monte_carlo_simulation — Price forecasting: conformal-calibrated risk by default (honest VaR/CVaR + Safe-Hold/Momentum grades