Financial MCP Servers & AI Agents

Ask one question before any feature list: who publishes the server. An official server, from the company whose data it serves, moves when the API moves and is somebody's job; a third-party wrapper breaks silently the next time an endpoint changes, and both sit in the registries with identical badges. After that, two things — the server is nearly always free while the data behind it is not, and every tool it exposes is injected into your context on every message, so a small tool surface is a running cost saved.

• Free, no API key, public source — SEC EDGAR MCP.
• A full open-source research platform with its own server — OpenBB.
• Local database, no vendor key — MaverickMCP.
• Fundamentals and prices for an agent — Financial Datasets.
• Institutional filings research — AlphaSense, if enterprise pricing is not an obstacle.

This listing is the products bought for this job. A data provider, tracker or backtester that also ships an MCP server beside its main product keeps its own category and is not here — those are collected, across every category, under tools with MCP. If what you need is an MCP endpoint for a vendor you have already chosen, start there instead.

Two jobs, and one of them is nearly empty

Infrastructure is a tool an agent can call: quotes, filings, fundamentals, options flow, order placement. You already have the agent; what you are choosing is the pipe. It is judged on coverage, licence, whether it is alive, and how much of your context window it eats.

Research copilot is the other end — the product is the answer, not the tool call. You ask about a company and it reads filings and calls and comes back with a paragraph.

Almost everything that survives a first filter here is infrastructure. That imbalance is not an accident of what we have got round to writing: the copilot half consolidated. The serious institutional product in that space, AlphaSense, is not in this listing because it is enterprise-only at roughly $10,000 to $20,000 a seat a year on an annual contract. Fintool, the credible startup challenger, was acquired by Microsoft in April 2026 and its team and technology went into Office. What is left in the open is research code from academic groups — FinRobot and FinGPT are papers with repositories attached, not tools you buy a seat on. Read that half of the listing as a survey of what people are building, and the infrastructure half as things you can actually install this afternoon.

Comparing across the two tells you nothing, because they fail differently. A bad data server returns nothing and you notice. A bad copilot returns a confident paragraph and you do not.

What actually decides the choice

Whose server is it. This is the first question and the one the directories hide. An official server is published by the company whose data it serves; a third-party server wraps someone else's API from the outside. The official kind now covers most of the large vendors — Alpha Vantage, Massive (the former Polygon.io), Alpaca, EODHD, Twelve Data, Unusual Whales, QuantConnect and OpenBB all publish their own, and many of them are in the market data APIs listing already, with the server as a feature rather than a separate product. A third-party wrapper is one person's side project; it breaks when the upstream API moves and there is nobody whose job it is to notice. Both kinds sit next to each other in every registry with identical badges. Alpha Vantage alone has an official server and at least three independent copies of it.

No upstream API means every server is a scrape. TradingView publishes no public data API, so every one of the at least six TradingView MCP servers on GitHub is built on the undocumented endpoints the site calls itself, or on a headless browser. They work until they do not, and nothing about their quality changes that. Where an official API exists, prefer a server that uses it.

The server is free; the data is not. Almost every vendor server refuses to start without an API key, so what you are really buying is a tier of the underlying subscription — and an agent calls tools in bursts of ten and twenty, which hits a per-minute rate limit in a way a script never did. Price the plan you will end up on, not the one the free key gives you. The servers that read a public source instead, such as SEC EDGAR, need no key and have no bill.

Stars record a launch; the commit log records now. This category churns harder than any other on the site — of the nine products worth naming eighteen months ago, two are gone and two turned into features of products we list elsewhere. The most-starred community server for a fundamentals API has over 2,300 stars and a last commit in June 2025. Check the last commit, the last release, and whether the open issues are answered, in that order.

AGPL is not background noise here. An MCP server is a network service, which is precisely the case AGPL section 13 covers, and the licence is unusually common in this corner — OpenBB's platform and its MCP extension are both AGPL-3.0, and so is SEC EDGAR MCP. Personal use on your own machine is unaffected. Exposing one to colleagues or customers is a decision for whoever owns your licensing policy, and it is cheaper to ask first. MIT and Apache-2.0 servers, such as MaverickMCP and QuantConnect's, carry none of this.

Ingested data or a pass-through. A proxy server owns nothing: it forwards your key to a vendor and the vendor's coverage, history and outage are yours. A server with its own store holds data locally, which means it works offline and costs nothing per call, but the freshness and the coverage are now the project's problem rather than a vendor's. MaverickMCP is the second kind — it keeps a local database and pulls prices through yfinance, so it needs no key and inherits every limitation of that source.

Tool count is a bill you pay on every message. Each tool definition is injected into the model's context before you have said anything. A server that exposes one tool per endpoint runs to fifty or more; a few of those installed together can consume tens of thousands of tokens per request and measurably degrade the model's choices. The vendors have noticed — Massive rewrote its server in March 2026 from 53 endpoint-shaped tools down to four composable ones, and OpenBB's server keeps the initial list short and lets the agent activate categories on demand. Prefer a small tool surface, and install fewer servers than you think you need.

Directory rankings in this space are advertising. The registries list SEO-farmed servers with no users beside real ones, and several of the "best MCP server for market data" comparison posts that rank well are published by one of the vendors being ranked — which is, predictably, the one that wins. Stars, a recent commit and a named publisher are cheap filters that remove most of it.

Where this category connects

Most of the servers worth installing are a feature of a data vendor rather than a product of their own, and the subscription behind the key is the real purchase — that is market data APIs. If the agent is meant to trade rather than read, the execution side lives in backtesting and algo frameworks, and reading a user's actual positions needs a brokerage aggregation API.