Skip to main content
Glama

🧟 FrankenFish — the second-hand electronics shop assistant stitched from spare parts

Kiroween 2026 · Frankenstein category · built with Kiro

FrankenFish stitches together pieces that were never meant to meet:

Body part

What it is

Role

🦴 Skeleton

goofish-cli (Apache-2.0) MCP server

Talks to Xianyu (闲鱼), China's largest second-hand marketplace, on the seller's own account

🧠 Brain

Any OpenAI-compatible LLM (domestic or overseas, user's own key)

Drafts buyer replies in Chinese

🫀 Heart

FrankenFish MCP server (this repo)

Price radar, pricing rules for used phones/laptops, reply guard rails

⚡ Lightning

Kiro specs, hooks and steering

How the creature was assembled

Who it is for: one-person second-hand electronics shops on Xianyu. The owner does sourcing, listing, customer service and after-sales alone. FrankenFish answers three questions every day:

  1. What should I price this used iPhone 15 Pro 256G 95新 at? → suggest_price

  2. What is the market doing for this model this week? → market_snapshot

  3. How do I answer this buyer who is haggling / asking about battery health? → draft_reply (never sends on its own)


Quick start (demo mode — no Xianyu account needed)

pip install -e .
frankenfish demo        # colourful walkthrough from bundled synthetic data

Expected output: market snapshot → price suggestion → three buyer reply drafts. Runs in under one second.

MCP server (Kiro / Claude Code / any MCP client)

{ "mcpServers": { "frankenfish": { "command": "frankenfish", "args": ["mcp"] } } }

Three tools are available:

Tool

Description

market_snapshot(model)

p25 / median / p75 of listing prices, outliers removed

suggest_price(model, condition, battery_health?, full_box?, has_warranty?)

list price (ends in 9) + floor price + basis explanation

draft_reply(buyer_message, list_price, floor_price)

classifies message, drafts Chinese reply, flags off-platform words, never auto-sends

Every tool returns {"ok": bool, "data": ..., "error": str|null, "source": str}.


Related MCP server: xianyu-mcp

Live mode (optional — seller's own account only)

  1. Install goofish-cli: pip install goofish-cli

  2. Authenticate: goofish auth login --qr

  3. Set the env var: FRANKENFISH_SOURCE=goofish

FrankenFish queries goofish-cli at most 3 times per minute and caches results for 1 hour. If goofish-cli is missing or not logged in, it falls back to demo data automatically.


LLM reply drafting (optional)

Set these env vars to enable LLM-drafted replies (falls back to rule-based drafts when unset):

FRANKENFISH_LLM_BASE_URL=https://api.openai.com/v1
FRANKENFISH_LLM_KEY=sk-...
FRANKENFISH_LLM_MODEL=gpt-4o-mini   # any OpenAI-compatible model

The same guard rails apply to LLM output: off-platform words are flagged, bargain offers below floor_price are clamped, and auto_send is always False.


Pricing rules

Input

Adjustment

Condition grade (全新/99新/95新/9新/8新/7新及以下)

×1.10 to ×0.62 of market median

Battery health < 85%

−5%

Full original box included

+2%

Valid Apple/manufacturer warranty remaining

+3%

Outliers (< 40% or > 200% of raw median)

excluded before statistics

list_price

rounded to nearest ¥10, then −¥1 (ends in 9)

floor_price

min(list_price, p25 × factor) × 0.97, capped at list_price


Tests

pip install -e ".[dev]"
PYTHONPATH=src python -m pytest -q

24 tests including Hypothesis property tests for every pricing rule (R1–R3) and reply guard rails (R4.1–R4.4).


Responsible use

  • Demo data in src/frankenfish/data/ is synthetic (every row has "synthetic": true).

  • Live mode reads a small number of on-demand searches on the seller's own account only.

  • No bulk scraping, no storing other users' data beyond a 1-hour in-process cache.

  • Replies are never sent automatically (auto_send is always False).

  • See docs/compliance.md for full details.

How Kiro was used

See .kiro/ (specs, steering, hooks) and docs/kiro-writeup.md.

License

MIT. Uses goofish-cli (Apache-2.0) as an optional external dependency — see NOTICE.


中文说明

FrankenFish(缝合鱼)是给闲鱼数码个体店主用的 AI 助手:行情雷达 + 智能定价 + 客服回复草稿,以 MCP 工具形式接入 Kiro、Claude Code 等。

默认演示模式用合成数据,不需要闲鱼账号;实盘模式只用卖家本人账号、按需少量查询(≤3次/分钟,缓存1小时),不批量爬取、不自动发消息。

可选 LLM 接入(OpenAI 兼容接口),未配置时降级为规则模板。每条定价建议都有 basis 字段解释每一步调整,议价回复永远不低于 floor_price。

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables automating buyer-side workflows on Xianyu (Goofish) by searching products, sending individual or batch price inquiries, and retrieving seller responses through MCP tools.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables searching Xianyu (Goofish) listings, fetching item details, taking keyword-monitor snapshots, and publishing secondhand products through MCP tools with local-session authentication.
      MIT