Disclaimer
Personal research and a technical experiment. Not investment advice, not a recommendation to buy or sell any security, and not affiliated with or endorsed by my employer.
Context
Most investing advice on the open internet is unattributable consensus. Reasonable on the surface, indistinguishable in the aggregate. The four investors who left the clearest records of their actual decision process, Warren Buffett, Peter Lynch, Benjamin Graham, and Charlie Munger, each applied a distinct lens that survives examination. They disagreed with each other, often. That disagreement is the value.
MCP Investment Mentors is a Model Context Protocol server that exposes four of those lenses as separate personas. Ask Buffett about a stock and you get a moat-and-management read. Ask Lynch about the same stock and you get a products-you-actually-use read. The mentors are not impressions. Each one is anchored to a corpus of letters, interviews, and published commentary, and instructed to refuse questions it would have refused in life.
What was built
An MCP server with four tools, one per mentor, each backed by a system prompt and a small toolkit of read-only API calls. The personas can fetch recent filings, basic financials, and a price chart. They cannot trade, recommend, or pretend to know things they would not have known.
The interesting work was in the refusal logic. A Buffett persona that gladly opines on early-stage SaaS is useless. The system prompts encode each investor's known no-go zones, and the server returns a structured refusal when a question lands inside one. Munger refuses anything that requires forecasting macro. Graham refuses anything below his margin-of- safety threshold without running the numbers. Lynch refuses on companies he has no plausible reason to understand. Buffett refuses on anything outside the circle of competence that the corpus actually documents.
A second design call: no consensus output. The server never blends four answers into one. If two mentors agree, the agreement is visible because each spoke separately. If they disagree, that disagreement is the deliverable.
What was learned
The MCP layer turned out to be the decisive choice. Once the personas were addressable as separate tools, integration with any MCP-capable client was trivial. The model client routes the question, gathers the four readings, and presents them side by side without further engineering.