Taobao Sourcing Assistant

A local, human-paced MCP server that removes the drudgery of sourcing products on Taobao/Tmall. You keep all judgment (search intuition, buy decisions, sending supplier messages); the tool drives a real Chrome window to extract — for every product — a price for every SKU variant, specs, images, and reviews linked to the variant bought, then tabulates it into a comparison spreadsheet. Ships with a Claude Skill (sourcing playbook) and Chinese supplier-message templates (drafted by Claude, sent manually by you).

Built on the QR-login + persistent-session approach of JeremyDong22/taobao_mcp, rebuilt as 6 FastMCP tools with embedded-data + DOM extraction (mtop interception kept as a fallback), per-SKU pricing, variant-linked reviews, xlsx export, and a captcha human-handoff.

What it does NOT do

No headless scraping, no proxy rotation, no captcha-solving service, no auto-messaging, no cloud. Not getting your account flagged is the priority, not speed.

Install (one time)

# from the project root
uv venv --python 3.12
uv pip install -e ".[dev]"

You need Google Chrome installed (the real app, not Chromium, not Comet). The launcher is pinned to it in config.toml. If Chrome lives somewhere non-standard, edit [browser] executable_path, or clear it ("") to let Playwright resolve the chrome channel.

Configure

Edit config.toml (defaults are sensible):

• [browser] executable_path — pinned Google Chrome binary (avoids launching Comet/other Chromium).

• [browser] user_data_dir — the persistent profile (your login lives here; gitignored).

• [pacing] — random delays + max_products_per_minute (keep it low).

• [limits] — max_reviews, review_pages.

• [output] dir — where xlsx + run.log land.

Run

.venv/bin/python server.py                                   # stdio MCP server
npx @modelcontextprotocol/inspector .venv/bin/python server.py   # interactive inspect

For Claude Desktop, register it as an MCP server pointing at the full venv python path and server.py (use absolute paths — /Volumes/...).

First-run login (once per session)

1. Call taobao_initialize_login (or just taobao_fetch_product — it auto-ensures login).

2. A visible Chrome window opens to the Taobao QR page.

3. Scan the QR with your Taobao app. The server polls and continues automatically.

4. The session persists in user_data_dir — restarts reuse it, no re-scan.

Tools

The Skill

skill/SKILL.md is the sourcing playbook (search → you pick → fetch → translate → summarize reviews → normalize price-per-unit → compare → export → flag risks). skill/supplier_templates.md has Chinese message templates — Claude drafts, you send via Wangwang.

Troubleshooting

• It launched Comet / the wrong browser — set [browser] executable_path to your Google Chrome binary (default: /Applications/Google Chrome.app/Contents/MacOS/Google Chrome).

• "login_required" / NotLoggedInError — run taobao_initialize_login and scan the QR; keep the window open.

• A slider/verification appeared — solve it yourself in the Chrome window; the tool pauses (human_action_required) and resumes. It logs to output/run.log.

• Screenshots/automation "page still loading" — the new detail page holds a connection open; this server uses embedded-data + DOM extraction (not screenshot-waits), so this only affects ad-hoc scripts.

• SelectorDriftError — Taobao changed its layout; patch the one file src/extract/selectors.py.

• Wrong price on a multi-model listing — the headline price is the cheapest model; always read the per-SKU price for the exact variant. 补贴后 prices may include a 国补 subsidy that needs a mainland ID — verify the real checkout price.

• Only a few reviews returned — deep review pagination is shallow (known limit); increase scrolling in src/extract/reviews.py if needed.

• Reset everything — delete user_data/chrome_profile/ and re-scan the QR.

Risks (don't hide these)

• Scraping Taobao violates its ToS; using your own logged-in account carries account-limitation risk. Keep volume low and human-paced.

• mtop endpoints / selectors drift — budget periodic maintenance (selectors are centralized).

Tests

.venv/bin/python -m pytest -q     # parsers, output, MCP contract, drift, evals