Resale Agent Skill Hub MCP Server

SkillHub — Smart Resale Agent

A collection of 8 Claude Code skills + an MCP Server implementing a Multi-Platform Smart Resale Agent: an AI-powered conversational workflow for C2C second-hand resale that reduces the barriers to listing and selling pre-owned goods.

v2: Now runs as a standard MCP server (JSON-RPC over stdio) that any MCP-compatible client can call — Claude Desktop, Cursor, VS Code, or custom web frontends.

Architecture

skillhub/
├── broker_core/                  # Shared engine (pip install -e .)
│   ├── session_manager.py        # Browser login → cookie persistence
│   ├── platform_client.py        # Authenticated platform API clients
│   ├── state_manager.py          # Listing lifecycle state (~/.broker/listings.json)
│   ├── scheduler.py              # launchd (macOS) / cron (Linux) registration
│   ├── audit_logger.py           # Append-only Transparency Log (~/.broker/audit.jsonl)
│   └── mcp_server.py             # ★ MCP Server — 7 tools via JSON-RPC
├── guardrails/                   # ★ Prompt framework (v2)
│   ├── Omni-Agent-Guardrails.yaml
│   └── README.md
├── skills/                       # 8 Claude Code skills
│   ├── personal-broker/          # Hub: orchestrates the 7-step pipeline
│   ├── broker-recognize/         # Step 1: Photo → product info
│   ├── broker-auth/              # Step 3: Bind platform accounts
│   ├── broker-price/             # Step 2: Authenticated price research
│   ├── broker-card/              # Step 4: Listing cards → auto-publish
│   ├── broker-schedule/          # Step 5: Repricing interval & cron/launchd
│   ├── broker-fuse/              # Step 6: Price Shield (floor price)
│   └── broker-delist/            # Step 7: Auto-detect sale → delist all
├── tests/                        # 41 tests
└── docs/                         # Setup guides for MCP clients

The 7-Step Pipeline

📷 broker-recognize  →  Photo → structured product info
🔐 broker-auth       →  Open browser → log in once → session persisted
🔍 broker-price      →  Search platforms with session → transparency report
🛡️ broker-fuse       →  ⚠️ GLOBAL INTERCEPTOR — set unbreachable floor price BEFORE publishing
📋 broker-card       →  Generate cards → validate floor → user confirms → auto-publish
⏰ broker-schedule   →  Register launchd/cron for daily repricing checks (floor-gated)
✅ broker-delist     →  Scheduler detects sale → auto-delist all platforms (no confirmation)

Each skill can also be invoked standalone (e.g., /broker-price for pricing only).

Prerequisites

ANTHROPIC_API_KEY (required)

broker-recognize uses the Anthropic API (Claude Vision) to analyze product photos. Set your API key:

export ANTHROPIC_API_KEY="sk-ant-..."

Without this key, photo recognition will fail.

Python 3.9+

Playwright (for platform auth)

Quick Start

Pick one path. Path A is recommended for most users.

Path A: Claude Desktop (recommended, 5 minutes)

Step 1: Install

git clone https://github.com/madguyevans-creator/skillhub.git
cd skillhub && pip install -e .
pip install playwright && playwright install chromium

Step 2: Get API key (one-time) Go to console.anthropic.com → API Keys → create key. Save it once:

echo 'sk-ant-your-key' > ~/.broker/api_key

Step 3: Configure Claude Desktop Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "skillhub": {
      "command": "python3",
      "args": ["/path/to/skillhub/broker_core/mcp_server.py"],
      "env": { "BROKER_MOCK_MODE": "true" }
    }
  }
}

Step 4: Restart Claude Desktop. Done.

Open Claude Desktop, say "帮我卖掉这双鞋" + upload photo. Claude automatically calls 7 MCP tools in the correct order.

Path B: Claude Code CLI

git clone https://github.com/madguyevans-creator/skillhub.git
cd skillhub && pip install -e .
cp -r skills/* ~/.claude/skills/

Use inside Claude Code:

• /personal-broker — full pipeline
• /broker-recognize — analyze product photo
• /broker-price — price research
• /broker-fuse — set floor price

Path C: Demo without any accounts (mock mode)

By default BROKER_MOCK_MODE=true. Everything works with realistic simulated data — no API key, no platform accounts needed. Photo recognition returns a demo Nike sneaker.

git clone https://github.com/madguyevans-creator/skillhub.git
cd skillhub && pip install -e .
# Done — you're in mock mode. Use Path A or B above.

Going real: turn off mock mode

export BROKER_MOCK_MODE=false
# Then bind platform accounts:
python3 skills/broker-auth/scripts/auth.py --all
# This opens a browser. Log into each platform once. Cookies persist.

Sandbox / Mock Mode

By default, BROKER_MOCK_MODE=true — all platform API calls (search, publish, delist) return realistic simulated data. No real HTTP requests are made. This allows:

• Demo without platform accounts: test the full pipeline with mock data
• Safe development: no risk of accidentally publishing or modifying real listings
• Anti-scraping avoidance: no requests that could trigger rate limits or CAPTCHAs

To use real platform APIs, set BROKER_MOCK_MODE=false in your environment. Requires platform accounts bound via broker-auth.

Transparency Log

Every AI-assisted decision is recorded to ~/.broker/audit.jsonl — an append-only, immutable JSONL file. You can always audit what the agent did and why.

What gets logged:

• Price research results (log_pricing)
• Price shield checks — blocked or passed (log_fuse_check)
• Listing publishes (log_publish)
• Auto-repricing events (log_repricing)
• Delist events (log_delist)
• Schedule registrations (log_schedule_registered)

How to view:

from broker_core import audit_logger
print(audit_logger.format_trail("item_abc123"))  # Human-readable timeline
trail = audit_logger.get_audit_trail("item_abc123")  # Raw JSON array

When every action is recorded with rationale, there is no "black box" — the user can always audit what happened and why.

Key Design Decisions

Decision	Rationale
Not web scraping	Uses user's own session cookies. Searching as a logged-in user, not crawling.
Bind once, reuse	Platform login happens once via browser. Session persisted until expiry (~30 days).
Sold median pricing	Recommended price = median of actually-sold comparables. Falls back to 5% below active median.
Price shield as global interceptor	Every price-mutating action (publish, repricing) validates against the floor. Blocked actions are logged to audit trail.
Auto-execute after confirm	Once user confirms card + floor + schedule, the system auto-publishes and auto-reprices without re-confirmation.
No confirmation on delist	When scheduler detects sale on any platform, auto-delist all others immediately.
launchd / cron	System-level scheduling, no daemon process needed.
Mock mode	BROKER_MOCK_MODE=true by default — safe demo without real accounts or anti-scraping risk.
Append-only Transparency Log	Immutable JSONL log of every decision. The user can always audit what happened.

License

MIT — see LICENSE