Shibui Finance
The financial grounding layer for Claude
10 followers
The financial grounding layer for Claude
10 followers
Stop Claude from hallucinating stock data. Free connector to 9,900+ US equities, fundamentals, technicals, and SEC filings. Screen, analyze, and backtest.
Free
Launch Team / Built With
AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
Maker Story: How Shibui Finance Started
The problem
I'm a freelance data engineer in Berlin. I build ETL pipelines, PostgreSQL databases, dbt transformations - that's my day job. A couple of years ago I decided I wanted to learn how to invest. Not day trade, not gamble on meme stocks. Actually understand how to evaluate companies.
I had no idea where to start. I didn't know what P/E meant. I'd never read a balance sheet. But I knew how to build data infrastructure, so that's what I did.
I started pulling stock data. Daily prices, quarterly financials, technical indicators. Built the pipeline, loaded it into PostgreSQL, ran dbt on top. Within a few months I had 31 million rows of price history going back to 1962, quarterly fundamentals from 1990, 56 technical indicators, daily valuations. A proper data warehouse.
Then I realized: having all the data in the world doesn't help if you don't know what to look for.
I had a database that could answer almost any question about US stock markets. I just didn't know the right questions yet. I ordered books. Tom Hougaard on trading psychology. Turner on day trading. Volman on price action. Started reading, started testing what they said against my data.
The turn
Around that time, Anthropic shipped MCP - Model Context Protocol. A way for Claude to talk directly to external tools and databases. I looked at what I'd built and thought: if I can point Claude at this database, I can just ask questions in English and get real answers back.
So I wrapped the database in an MCP server and connected it to Claude. The first time I asked "show me companies where profit margin exceeded 15% every single quarter for the past 5 years" and got back a clean, correct answer in seconds - that was the moment. That question is impossible on Finviz. It's awkward on Bloomberg. It's one sentence here.
How it evolved
What started as a personal learning project kept getting more capable because I kept running into things I couldn't do elsewhere.
First phase: just get the data in. Daily OHLCV, quarterly statements, annual statements. Basic pipeline work, nothing unusual. I spent weeks cleaning symbol mappings and handling corporate actions. Nobody sees this work, but it's the difference between a database that gives wrong answers and one that doesn't.
Second phase: derived metrics. Raw financial statements don't have P/E ratios or free cash flow yield. I built a dbt layer that computes daily valuations, derived fundamentals, Piotroski F-Scores. Pre-computed 56 technical indicators so Claude doesn't have to calculate RSI from scratch every time someone asks.
Third phase: the MCP server. This is where the product emerged. I gave Claude direct SQL access to the database through a structured tool. Not an API wrapper that calls one endpoint per question - actual SQL against a real database. That distinction matters. An API wrapper can tell you Apple's P/E. It can't find every company where P/E stayed below 15 for 20 consecutive quarters, because that requires a cross-table join with window functions across thousands of symbols. SQL can. API wrappers physically can't.
Fourth phase: real users showed up. I shipped it free, listed it on MCP directories, and waited. Within weeks, an anonymous user started running halal stock screening queries using AAOIFI criteria. He came back daily for two months. Over 100 sessions. He built a 6-factor scoring system entirely through conversation - no code, no spreadsheets. He calls it the "IQ Score." I still don't know his name.
Another user started pulling SEC filings data. Someone else ran backtests in Korean. A forensic accountant tested 8-K filing queries. None of them needed to learn SQL or Python. They just asked questions.
Fifth phase: SEC filings. Users wanted to know about insider buying, institutional ownership changes, 13D activist positions. I added a feed that pulls SEC EDGAR filings every 5 minutes - Forms 3, 4, 5, Schedule 13D/G. Now you can ask "which insiders bought stock in the last week at companies where the stock is down 20% from highs" and get a real answer.
What I learned building it
The hardest part was not the engineering. It was restraining myself from engineering.
My instinct was to keep building infrastructure - add more data sources, optimize queries, clean up the schema. Meanwhile, a real user built a portfolio management system on my data, and I broke his workflow with a schema change because I had no way to contact him. No email list, no changelog, no notification system. Classic builder trap: I was optimizing the engine while a driver was mid-race.
The other thing I learned: every stock screener shows you today's snapshot. Current P/E. Latest quarter's margin. That's useful but it answers the wrong question. The interesting questions are almost always temporal. "Has this company maintained high margins consistently?" "Did revenue grow every single year through COVID?" "What happens to stock prices 90 days after the first earnings miss following a long beat streak?" These are one-sentence questions in Shibui. They're impossible on traditional screeners and expensive on institutional tools.
What it is today
Shibui Finance is a free MCP server at mcp.shibui.finance. Connect it to Claude, ask a question about US stocks, get real data back. About 10,000 symbols on NYSE and NASDAQ. 64 years of daily prices. Quarterly financials from 1990. Daily valuations from 1993. 56 technical indicators. SEC filing data updated every 5 minutes.
It's not Bloomberg. No intraday data, no options, no real-time feed, roughly 1-day lag on prices. For screening, backtesting, company analysis, and temporal pattern detection, it has more depth than anything else at this price point. The price point is zero.
Built by one person, maintained by one person, used by people I'll probably never meet, answering questions I never thought to ask.