Best MCPs for Finance

July 12, 2026

Your investment committee wants an answer on Nvidia by Friday. Add, trim, or pass. The analyst assigned to it will spend three days doing what analysts have always done: pulling the 10-K, skimming eight quarters of earnings calls, reconciling consensus numbers, checking what insiders are selling, and formatting all of it into a memo.

Almost none of that is analysis. It is retrieval. And retrieval is the part machines just got very good at.

We made the argument in Alpha Runs on Context: the model is a commodity, the edge is what you feed it. The Model Context Protocol is how the feeding happens. This post is the practical follow-up: the stack we would stand up on a hedge fund's research desk Monday, and why each layer earns its slot.

The Thirty-Second Primer

MCP is the open standard Anthropic released in November 2024 for connecting AI systems to external data and tools: "secure, two-way connections between their data sources and AI-powered tools." It won. ChatGPT, Gemini, Microsoft Copilot, Cursor, and VS Code all speak it, and in December 2025 Anthropic donated the protocol to the Agentic AI Foundation under the Linux Foundation, with more than 10,000 active public MCP servers in the wild.

For an investment team, the translation is simple. An MCP server turns a data vendor into a tool your AI agent, a model that can take actions rather than just answer questions, can call directly. The integration work collapses to a URL and a sign-in.

The diligence does not collapse. Easy to connect is not the same as cleared to connect. Hold that thought; we will come back to it.

The Stack

Breadth First: Alpha Vantage

Alpha Vantage's MCP server is first-party and official, offered as a hosted endpoint with OAuth or as a local install. One connector covers fundamentals, news with sentiment scoring, earnings-call transcripts, insider and institutional activity, and prices across every asset class an equity desk touches. For the Nvidia question it is the first pass: quarterly numbers, street sentiment, and insider selling in one conversation instead of five browser tabs.

The catch: it is an aggregator behind a hosted endpoint, which means your queries leave your perimeter, and the free tier is rate-limited well below institutional volume. Budget for a paid key, and reconcile anything memo-bound against the primary filings on EDGAR. That reconciliation is not busywork; it is what keeps an agent's memo defensible.

Synthesis: Octagon

Where Alpha Vantage returns data, Octagon returns research. Its agents work SEC filings across what the vendor reports as 8,000+ public companies, ten years of earnings-call transcripts and financial metrics, and market data for 10,000+ tickers. One agent runs deep research; another, unusually, tracks Kalshi prediction markets.

The catch: those coverage figures are Octagon's own, we could not verify them independently, and the vendor's fine print dates filings from 2018 onward. It runs hosted-only, behind an API key, from a young company without a public track record. Treat its output as a fast associate's draft: genuinely useful, never citable until you have checked it against the filings themselves.

The Institutional Route: Claude for Financial Services

FactSet, Morningstar, PitchBook, S&P Global, and Daloopa do not sell public MCP servers you can wire up yourself. If you want that data agent-ready today, the route is Claude for Financial Services, which ships pre-built MCP connectors to all five.

That buys speed. It also buys a dependency: the entire route runs through a single model vendor's enterprise offering, with access and pricing negotiated rather than self-serve. You are standardizing on Anthropic, not just on MCP, which is the opposite of the assemble-it-yourself control this post argues for. For a fund that already licenses these providers and wants results this quarter, the trade is often worth it. Make it with eyes open.

Where the Output Lands: Anchor

Three picks move data in. The fourth is not a data source at all, and that is the point: something has to catch what comes out, and today most agent output dies in one analyst's chat scroll, gone by the next meeting. A shared drive fixes storage but not format; an agent's interactive HTML report does not belong in a folder of spreadsheets.

Anchor is one answer: a shared workspace with its own MCP server, one URL that any MCP-compatible agent, Claude or otherwise, writes its output into, and the rest of the team opens in a browser. The Nvidia memo, the comp table, the charts land somewhere the PM can pull up in the Friday meeting. The catch: it is a young product, and it would be holding your live investment theses, which makes it a data-residency and retention decision, not just a convenience. Notion or SharePoint close the same loop with more friction and a worse fit for agent output. We think the loop matters more than the logo: research that ends in a shared artifact instead of a private transcript is research the committee can act on.

The Bloomberg Question

Every conversation about financial data ends at the same place: what about Bloomberg?

As of this writing, there is no public, customer-facing Bloomberg MCP server. Bloomberg has embraced MCP internally, building the middleware layer (authentication, authorization, rate limiting, metering, AI guardrails) to make agentic AI production-safe inside its own walls. Its agentic features surface inside the Terminal, not as a server you can point your own agent at. The GitHub repositories named "Bloomberg MCP" are unaffiliated community projects that require your own Terminal license and say so in their READMEs.

J.P. Morgan is similar in shape: its publicized generative-AI effort is LLM Suite, an internal platform for employees. We found no public J.P. Morgan MCP server.

The pattern is worth reading correctly. The incumbents are keeping their data inside their own products while the open MCP ecosystem grows around them. The stack above is not a compromise while you wait for Bloomberg. It is the current state of the art for agent-native research.

This Is Already Production

Bridgewater's AIA Labs built an Investment Analyst Assistant on Claude that generates Python, builds visualizations, and iterates complex financial analysis "with the precision of a junior analyst," in the words of its CTO. Balyasny reports that roughly 95% of its ~180 investment teams actively use its internal AI research platform, with agents synthesizing tens of thousands of filings, research notes, and earnings documents, compressing days of deep research into hours. Both are vendor-published case studies, and Balyasny's stack runs on OpenAI rather than MCP. But the direction is unambiguous: the largest funds have already re-tooled research around agents. The open question is only whether you assemble the stack yourself or wait.

Adopt Like a Regulated Firm

There is a tension in this post. We handed you four servers, then told you that easy to connect is not cleared to connect. Both are true. You do not deploy all four on Monday; you deploy the coverage you already pay for, behind scoped credentials and an allowlisted registry. The stack is the edge. The ungoverned stack is the incident.

The industry agrees on where the risk sits. In a December 2025 survey by Stacklok, an MCP-security vendor, 58% of the 100 senior financial-services technical leaders polled named security as the top barrier to MCP adoption, ahead of data quality and legacy integration. The same survey found 61% of financial firms enforce regulatory data boundaries on AI systems, against 41% across all industries. The named concerns are the right ones: prompt injection (a hostile instruction smuggled inside data an agent reads), tool poisoning (a compromised server handing your agent booby-trapped tools), data leakage across organizational boundaries, and unvetted servers appearing on analyst laptops.

For a regulated manager the checklist is not generic. Books-and-records obligations do not care that the memo came from an agent: retention and reproducibility decide where output is allowed to land. Supervision means a human signs the memo, not the model. And your information barriers do not dissolve because the leak now travels through a tool output. Map each connector to the obligation it touches before it reaches a laptop. We wrote the playbook in Governing MCP Servers, Skills, and Plugins at Enterprise Scale.

The Edge Is the Stack

Back to Nvidia. The analyst who answers the committee first is not smarter than yours. They are better plumbed: retrieval, synthesis, and verification wired into one governed loop, with the finished memo landing where the committee can open it.

The AI you can buy is the AI everyone has. The stack you assemble, govern, and own is the edge.