mgc-blackbox 1.4.3

A trusted boundary for intent‑safe AI execution.

MirginCipher Blackbox (MGC) — Encrypted AI Agent Execution Layer

A secure local execution layer for AI agents — encrypted storage, sealed scripts, zero plaintext leakage.
Protect API keys, credentials, and scripts from AI agents with AES‑256 + RSA hybrid encryption and a Cython‑compiled secure core.

---

What is MGC Blackbox?

MirginCipher Blackbox (MGC) is a Local Encrypted Execution Layer designed to protect sensitive human intent and enable secure, deterministic AI execution.
It provides a trusted device‑level encrypted boundary for agents — MGC is not an agent itself.

MGC ensures:

• Sensitive data never leaves the device
• AI agents cannot access plaintext
• Scripts execute inside a sealed, encrypted environment
• Cross‑node execution is possible without exposing code

---

Why MGC?

• 🔐 End‑to‑End Encrypted Storage
  AES‑256 encrypted vault for API keys, credentials, configs — never exposed to AI agents or external systems.

• 🧱 Local‑First Security Boundary
  All execution and decryption happen on‑device. No cloud dependency, no plaintext leakage, no telemetry.

• 🧩 Sealed Script Execution (Unique)
  Convert scripts into unreadable execution capsules.
  Only trusted nodes can decrypt & run them — even the sender cannot read sealed scripts.

• ⚡ Deterministic Local Execution
  Stable, reproducible behavior across macOS / Linux / Windows with a Cython‑compiled secure core.

• 🛠️ Native MCP / Skill Integration
  Exposes mgc_save / mgc_get / mgc_list / mgc_seal / mgc_open_webui as standard MCP tools.
  Works out‑of‑the‑box with Copilot, Claude, Trae, IDE Agents.

• 🔄 Zero Integration Cost
  Any MCP‑compatible agent can immediately use MGC as its secure execution backend — no SDK, no custom code.

• 🛡️ Designed for AI Agent Security
  Protects human intent, prevents agent overreach, and enforces strict execution boundaries.

---

Use Cases

1. Protect API Keys & Credentials from AI Agents

Store secrets encrypted. Agents can use them, but never see plaintext.

2. Secure Local Automation

Run Python / Shell / Node scripts locally without exposing sensitive data to AI logs or cloud systems.

3. Sealed Script Distribution

Share scripts with collaborators or devices without exposing source code — they can execute but cannot read.

4. Cross‑Node Execution

Send sealed scripts to trusted remote nodes:

• Sender cannot read the sealed content
• Recipient cannot read the sealed content
• Only the target node can decrypt and execute

Ideal for enterprise automation, multi‑node collaboration, and privacy‑sensitive workflows.

5. Local‑First AI Agent Security Boundary

Provides a local security layer for Copilot / Claude / Trae / IDE Agents:

• Local encrypted storage
• Local execution
• Local permission control
• No cloud dependency

6. Privacy‑Preserving AI Workflows

Enables financial automation, personal data processing, and enterprise internal workflows with privacy protection.

---

Architecture

---

Crypto Layer & Performance

MGC uses a hybrid cryptographic design:

• AES‑256‑GCM — bulk data encryption (vault items, sealed script payloads)
• RSA‑2048/4096 — key encapsulation, node authorization, cross‑node execution rights

The crypto_layer is Cython‑compiled to:

• Improve AES and especially RSA performance
• Reduce Python‑level overhead for large integer arithmetic
• Provide a sealed, tamper‑resistant execution boundary
• Prevent monkey‑patching and unauthorized modification
• Maintain deterministic behavior across nodes

Security does not rely on code being hidden.
We rely on standard cryptographic primitives and a clear threat model.
Compilation reduces attack surface and improves performance — not “security through obscurity”.

---

Features

• Local encrypted storage
  Sensitive data is encrypted and never uploaded to the cloud.

• Encrypted execution
  Scripts run inside the encrypted boundary; plaintext is never exposed to AI or external systems.

• Store‑once authorization
  Items can be reused within the same device environment without repeated confirmation.

• Environment migration
  If hardware changes, access can be restored using a user‑defined migration key.

• Cross‑agent availability
  Any agent platform supporting Skills / MCP can integrate with MGC with zero additional development.

• Cross‑platform support
  Distributed as a Python package with security‑critical components compiled via Cython.

• No delete function
  MGC treats all stored info as user assets.
  To delete: use WebUI → Database Audit → manually delete via DB Browser.

• Script Sealing (Cross‑Node Execution Rights)
  MGC can seal scripts into non‑readable execution capsules:

  • Ownership remains with the user
  • Execution rights can be granted to trusted external nodes
  • Only the target node can decrypt & execute
  • Sender cannot read sealed script contents
    Enables secure cross‑node execution without plaintext exposure.

---

Quick Start

1. Install

pip install mgc-blackbox

2. Start Service

mgc

Launches the encrypted execution layer and WebUI.

Default WebUI port: 57218
If occupied, MGC automatically decrements (57217, 57216, …).

3. Open WebUI

http://127.0.0.1:57218

4. Store a Secret

from mgc import save

save("openai_key", "sk-xxxx")

Or via WebUI.

5. Execute Scripts Securely

Scripts run inside MGC's encrypted boundary. Result is notified only (success/failure) —
output goes to external destinations (files, emails, webhooks, etc.), not returned to the caller.

---

Example: Save & Retrieve Secrets

from mgc import save, get

save("openai_key", "sk-xxxx")
print(get("openai_key"))

---

MCP Integration

MGC exposes a local MCP tools interface:

• mgc_save
• mgc_get
• mgc_list
• mgc_seal
• mgc_open_webui

Compatible with:

• Copilot Agent
• Claude Agent
• Trae Agent
• IDE Agents
• Custom Agents

MCP configuration file:
mcp_config.json (auto‑generated on installation)

---

Usage Overview

1. Through AI agents (Skills / MCP)

Agents can:

• Store sensitive information
• Retrieve encrypted items
• Execute stored scripts
• Seal scripts for trusted nodes (external / local)
• All without unauthorized plaintext access

2. Through system scripts (REST API)

External scripts can fetch encrypted items at runtime.
Plaintext is never exposed to AI logs or external systems.

For detailed usage:
MGC_GUIDE.md

---

Security Model

• All data remains local
• No cloud upload
• No plaintext logging
• Deterministic execution
• User‑controlled authorization
• Protection Mode for high‑security environments
• Minimal network usage (only version & health checks)

For safety details:
docs/user_notice.md

---

AI Skill Specification

For AI behavior boundaries and tool definitions:
docs/skill_spec.md

---

Authorization

Integration into any third‑party products or AI agents is free of charge,
but requires official authorization to ensure ecosystem integrity and security.

For authorization requests:
[e-mail]

---

License

See the LICENSE file for full terms.

© 2026 MirginCipher Team. All rights reserved.

---