Key points
- Wall Street tooling meets Chinese infrastructure: Anthropic's open-source financial-services stack (41 Skills, 38 Commands, 11 MCP connectors) targets US data terminals. Three community forks rewire the same plugin architecture for China: AKShare (ccq1), iFinD (NBreeze-Eric), and Qichacha (duhu2000).
- File-driven, zero-build design: Every plugin is just
plugin.json,.mcp.json,commands/*.md, andskills/*/SKILL.md— Markdown and JSON only. Analysts can edit SKILL.md to inject their own valuation conventions. - Three localization routes, three trade-offs:
- AKShare route: 42 tools, no API key, broadest coverage, suited to junior sell-side researchers.
- iFinD route: 4 MCPs over Tonghuashun's paid terminal, higher data fidelity, native Excel output.
- Qichacha route: 2 plugins (17 Skills), focused on KYB, risk-signal scanning, and IC Memo for banks, trusts, and PE/VC.
- 14 Commands cover the core workflow:
/comps,/dcf,/3-statements,/competitive-analysis,/earnings,/initiate,/screen,/morning-note,/sector,/thesis,/ipo,/merger,/pitch,/bond. The DCF skill defaults to Chinese 10-year sovereign yield as Rf, A-share historical risk premium (≈4–5%), and CSI 300 as the beta benchmark. - Deep localization beyond translation:
- Accounting switched from US GAAP/IFRS to CAS, with explicit handling of
deducted-non-recurring net profit. - Industry taxonomy moved from GICS to Shenwan (31 Level-1 / 134 Level-2 sectors).
- Trading rules: T+1 instead of T+0, with 10% / 20% / 30% price-limit bands per board.
- IPO logic auto-compares STAR Market, ChiNext, Main Board, and BSE thresholds.
- Bond coverage spans enterprise bonds, short-term financing, MTNs, PPNs, ABS.
- Quantified productivity gains: Research-report cycles drop from 6–9 hours to 1–2 hours; pitch books from 3–4 days to half a day; KYB onboarding from 2–3 hours to ~15 minutes.
- Real risks: AKShare data is delayed 15 min–1 day and can diverge from Wind/iFinD; models can misclassify Shenwan industries, confuse attributable vs. reported net profit, or fabricate WACC assumptions; CSRC compliance bars AI from issuing buy/sell ratings, and any client-facing service may need algorithm filing. All AI output must be reviewed by licensed analysts.
- Setup commands are included for each fork: AKShare via
git clone https://github.com/ccq1/cn-financial-services-plugins.gitplusclaude plugin install; iFinD via WCP Bearer Token insettings.local.json; Qichacha via the project's one-line installer script.
Why this matters
The forks illustrate a broader pattern: US AI stacks ship with US data plumbing, and Chinese developers don't wait for localization — they fork the skill layer and swap in domestic MCPs. The next milestones likely include a Wind official MCP, exchange-direct APIs, and regulator-grade AI audit tooling.