Multi-Agent Safe-Cooperation (Single Device)

One-Line Description

A single-device multi-Agent task chain collaboration methodology based on MGC. Through Master Agent orchestration, Script Agent scripting, and Executor Agent execution, achieves zero-exposure security collaboration for sensitive resources.

Problem Solved

Core Concepts

MGC's Role in Collaboration

MGC is the local sensitive resource hosting and execution layer (skill_spec.md §1):

1. User stores keys and scripts via WebUI (browser or mgc_open_webui)
2. Script Agent writes scripts and stores them in MGC via mgc_save (MGC 1.4.10 auto-parses argparse defaults into ext02)
3. Master Agent orchestrates the task chain, specifies call timing; can execute directly when sub-Agents fail or with user authorization
4. Executor Agent only calls mgc_run, never touches keys or scripts

Four Roles

Note: mgc_get retrieves plaintext and is a sensitive operation requiring user authorization; Script Agent typically does not need it.

Collaboration Flow

User ──Store Keys/Scripts (WebUI)──> MGC
                │
                ▼
User ──Issue Command──> Master Agent
                │
                ├──Decompose Task Chain
                │
                ▼
        ┌───────┴───────┐
        ▼               ▼
   Script Agent      Executor Agent
   (Write Scripts)  (Execute Tasks)
        │               │
        ▼               ▼
   mgc_save         mgc_run
        │               │
        └───────┬───────┘
                ▼
           MGC Execution
        (Return pid+status)

Quick Start

Prerequisites

1. Install MGC v1.4.9+ (1.4.10 recommended): pip install mgc-blackbox>=1.4.9
2. Start MGC: mgc (API port 57219, WebUI port 57218)
3. Configure MCP (add to AI client):

   {
     "mcpServers": {
       "mgc-blackbox": {
         "command": "mgc",
         "args": ["--mcp"],
         "env": { "PYTHONIOENCODING": "utf-8" }
       }
     }
   }
   

4. Store sensitive resources via WebUI or mgc_open_webui: API keys, database credentials, business scripts
5. Load prompt templates into each Agent

Sandbox mode (1.4.9+): When running inside a sandbox Agent (Trae Work / Workbuddy), install MGC in the system environment; otherwise MCP operations may be limited — in that case, call FastAPI directly.

Step 1: Configure Master Agent

Load prompts/master_agent.md into the Master Agent's system prompt.

Step 2: Configure Script Agent

Load prompts/script_agent.md.

Step 3: Configure Executor Agent

Load prompts/executor_agent.md.

Step 4: Start Collaboration

After user issues a command, the Master Agent automatically decomposes the task chain and coordinates sub-Agents.

Prompt Templates

MGC Tool Usage

mgc_save (Used by Script Agent)

# Store script (ext01 required; ext02 optional — MGC 1.4.10 auto-parses argparse defaults)
mgc_save(
    info_type="script",
    info_owner="Data Query Script",
    ext01="python",
    content="""import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--query', default='default')
args = parser.parse_args()
print(args.query)"""
)

mgc_find (1.4.10 new, fuzzy search)

Used to fuzzy-match entries by info_owner / diff fields. Never returns content plaintext.

# Find scripts whose owner contains "query"
scripts = mgc_find(info_owner="query", match_mode="substring", limit=50)
# match_mode: substring (%x%) / prefix (x%) / suffix (%x) / exact (x)

mgc_run (Used by Executor Agent, 1.4.7+ blackbox execution)

⚠️ Important (1.4.10 contract): ext02 must be a JSON array string, matching the script's argparse argv list. Dict-style {"k":"v"} is no longer accepted and triggers HTTP 422.

Auto-parsing: After Script Agent stores a script, MGC auto-fills ext02 from argparse defaults. Executor Agent can omit ext02 to use defaults.

import json

# Recommended: build JSON array string with json.dumps
params = ["--query", "2026-08-08"]
result = mgc_run(
    info_type="script",              # required, type is script
    info_owner="Data Query Script",  # required, script name
    diff_1="v1",                     # required to uniquely identify when multiple entries share the owner
    ext02=json.dumps(params)         # optional JSON array string: '["--query", "2026-08-08"]'
)
# Returns: {"pid": 12345, "status": "started"}
# Note: mgc_run does not return stdout; for results, the script should write to a file and print the path

mgc_list (All Agents, exact match)

Note: mgc_list only lists entry metadata. Since 1.4.10 prefer mgc_find for fuzzy lookup.

scripts = mgc_list(info_type="script")

mgc_open_webui (User)

Used to have AI open MGC WebUI for users to store sensitive resources.

mgc_open_webui()

ext02 Auto-Parsing (1.4.10 Important Change)

Since MGC 1.4.10, when storing a script it auto-parses argparse literal defaults and stores them as a JSON array string in ext02:

# After Script Agent's mgc_save, MGC auto-fills ext02:
# '["--start", "2026-08-08", "--verbose"]'

# Executor Agent can omit ext02 to use defaults
mgc_run(info_owner="my_script", diff_1="my_script")

# Or override at runtime
mgc_run(
    info_owner="my_script",
    diff_1="my_script",
    ext02='["--start", "2026-12-25"]'
)

Dynamic defaults are not supported (e.g. datetime.now(), os.path.expanduser("~"), f-strings). MGC returns dynamic_args_detected warning and asks you to set ext02 manually. Script Agent should always use literal defaults.

Security Boundaries

What This Skill Provides

• Collaboration methodology framework
• Prompt templates
• MGC tool usage specifications

What This Skill Does NOT Provide

• MGC permission enforcement (relies on prompt constraints)
• Automatic key management
• Agent identity authentication

⚠️ Important Reminders

1. All sensitive resources must be stored by users via MGC WebUI
2. Executor Agent can only call mgc_run; must never read script content
3. Master Agent must clearly inform sub-Agents of call timing and script location
4. After sensitive operation results return, Master Agent is responsible for updating task status and coordinating sub-Agent collaboration
5. ext02 must be a JSON array string (e.g. ["--flag","value"]); never use dict-style

FAQ

Related Resources

• MGC Core Repository: https://github.com/zkeviny/MGC-Blackbox
• MGC Installation: pip install mgc-blackbox>=1.4.9
• WebUI: http://127.0.0.1:57218
• API: http://127.0.0.1:57219
• MGC API Auth Token: ~/.mgc/database/mgc_black_box/.mgc_token
• Feedback: mirgincipher@outlook.com

Changelog

v1.1.0

• Upgraded to adapt to MGC 1.4.10
• Fixed mgc_run ext02 signature: must be a JSON array string
• Added mgc_find fuzzy search tool
• Documented ext02 auto-parsing (1.4.10)
• Documented that mgc_run returns only pid+status (no stdout)
• Added 1.4.9 sandbox mode note
• Updated MGC main skill doc reference to WebUI "MGC Skills" button

v1.0.0

• Initial release
• Provides Master / Script / Executor Agent prompt templates
• Includes complete collaboration flow documentation