Two years ago, getting an AI assistant to answer a real research question meant pasting a document into the chat window or writing glue code that fetched data, trimmed it to fit a context window, and shipped it into a prompt. Both broke constantly. The paste was stale the moment you made it, and the glue code had to be rewritten every time a provider changed a response shape.

The Model Context Protocol changed that. MCP is a standard way to give an assistant like Claude direct, authenticated access to a data source. The provider runs a server that describes what it can do in a machine-readable way; your client connects once, and from then on the model can call those tools itself, with your credentials, mid-conversation. If you have never set one up, our explainer on what an earnings call MCP server is covers the mechanics.

The bottleneck has moved. It is no longer "can the model see my data" but "which servers do I connect, and which one is actually best at the job I need." What follows is a ranked list of eight, with honest notes on where each one stops being the right tool. MCP moves fast — tool counts and plan details change, so treat specifics here as a snapshot and check the provider's docs before you commit.

1. earningscalls.dev

Best for: what management and analysts actually said on the call, searchable across the whole market.

Our server lives at https://earningscalls.dev/mcp and exposes 18 tools over a corpus of 253,000+ earnings call transcripts covering 12,799 companies across 175 exchanges, from 2020 to today, broken into 11.99M speaker-tagged segments. The tools are built around how transcript research is actually done: list a company's call history, pull one transcript in full or as components, fetch speaker segments filtered by role so you can read only the executives or only the analysts, run full-text search across the entire corpus or inside a single ticker, and check the upcoming earnings calendar. It connects from Claude Desktop, claude.ai, Claude Code, Cursor and any other spec-compliant MCP client.

The limitation is deliberate and worth stating plainly: we do not do prices, we do not do fundamentals, and we cannot place a trade. If you ask our server what NVIDIA's forward P/E is, you will get nothing useful. We are the deepest option for earnings call narrative, and nothing else. That is why this list has seven more entries.

2. Polygon.io MCP

Best for: serious market data — bars, trades, quotes and options across asset classes.

Polygon maintains an official MCP server that exposes its REST surface as tools: aggregates and bars, trades and quotes, market snapshots, ticker details, corporate actions like dividends and splits, news, and reference fundamentals, across stocks, options, forex and crypto. You install it locally and point it at your Polygon API key. For anyone doing price-based work — expected moves, drawdown studies, event windows around an earnings date — this is the most rigorous data source on the list.

The limitation is entitlements. What the tools return depends entirely on which Polygon plan you are on, and the gap between the free tier and a real-time options subscription is large. An assistant calling a tool you are not entitled to will get an error rather than data, which is a confusing failure mode if you have not read your own plan carefully.

3. Alpha Vantage MCP

Best for: breadth on a free key — equities, FX, commodities, macro and technical indicators in one connection.

Alpha Vantage runs an official hosted MCP server, which means you connect a URL with your API key rather than installing anything. The coverage is unusually wide for a single endpoint: stock quotes and history, foreign exchange, cryptocurrencies, commodities, economic indicators, technical indicators, and a news-and-sentiment feed. If you want one server that can answer "what did the dollar do against the yen last quarter" and "what is the current CPI print" without adding two more connections, this is it.

The limitation is throughput. Free-tier keys carry daily request caps that are easy to exhaust — an assistant looping across thirty tickers can burn a day's quota in one conversation. Good breadth server, poor scanning server unless you are paying for it.

4. Financial Modeling Prep MCP

Best for: one connection that covers fundamentals, statements, ratios and market data together.

FMP's data is wrapped by an actively maintained MCP server with a very large tool surface — company profiles, the three financial statements, ratios, analyst estimates and ratings, indices, ETFs, commodities, crypto, forex and technical indicators, spread across two dozen categories. Usefully, the server supports loading a subset of toolsets rather than everything at once, which matters more than it sounds.

That is also the limitation. A server that exposes hundreds of tools can degrade the assistant's ability to pick the right one, and you will get better results by loading only the toolsets you actually need. Data quality also varies by endpoint more than it does at a pure market-data vendor, so verify anything unusual against a second source before you act on it.

5. Financial Datasets MCP

Best for: clean, agent-friendly fundamentals and filings when you are building rather than browsing.

Financial Datasets runs a hosted MCP server aimed squarely at AI workflows. Income statements, balance sheets, cash flow statements, prices, news and SEC filings come back in shapes that are easy for a model to reason over without a lot of unit-wrangling. Interactive clients sign in directly; scripts and production agents use an API key against the API endpoint. The free tier is small but real, which makes it a low-friction way to test whether an agent design works before paying for anything.

The limitation is coverage. It is narrower than the large incumbents, weighted toward US listings, and the free tier's daily request allowance will not survive a serious backtest or a market-wide scan. Treat it as a well-designed fundamentals source for agents, not as a replacement for a full data vendor.

6. Quant Investing MCP

Best for: running a quantitative screen and backtesting it in the same conversation.

This one is a different shape from the rest. Quant Investing's server gives Claude access to their live screener over a global universe of roughly 24,000 companies, plus their newsletter archives, your own saved screens and watchlists, and a backtest engine — so you can ask for a value screen with specific factor constraints, see the names, and immediately test how that ruleset performed historically, without leaving the chat. Very little else on this list closes that loop.

Two limitations. It sits behind a specific paid subscription tier, so it is not something you casually add for one question. And there is a daily tool-call quota with backtests counted at a heavier weight — sensible design, but an assistant iterating enthusiastically on a screen hits the ceiling fast.

7. Alpaca MCP

Best for: actually executing — an assistant that can place orders in a paper or live account.

Alpaca's official MCP server is the one that crosses from research into action. It maps to the trading API: quotes and bars, positions and account state, and order placement across market, limit, stop and stop-limit types for stocks, ETFs, options and crypto. It defaults to paper trading, which makes it a genuinely good sandbox — you can have an assistant translate a thesis into an order ticket and watch what it does without any money at risk. For execution, it beats everything else here, including us, without qualification.

The limitation is the obvious one, and it is not a criticism of the software. Flipping the configuration to a live account means a language model can move real money, and the guardrails are yours to build. Keep it on paper unless you have a specific reason not to, and review every order on the broker's own dashboard rather than trusting the chat transcript.

8. Interactive Brokers MCP

Best for: reasoning over your actual portfolio — positions, balances, P&L and risk exposures.

IBKR opened a hosted MCP endpoint that you authorize with your own account through their login flow. Once connected, an assistant can see positions and cash balances, margin availability, realized and unrealized P&L, historical transactions, option chains and risk exposures, and can draft trade instructions for you. By design, those instructions do not execute on their own. If your question is "given what I already own, how exposed am I to this theme," no generic data server can answer it and this one can.

The limitation is scope. It is bound to your authorized account, so it is a portfolio tool rather than a market-data source — it will not help you research a company you do not hold. And you need an Interactive Brokers account in the first place, which is a higher bar than an API key.

How to combine them

Nobody needs eight. The realistic setup is two or three servers running side by side, chosen so their jobs do not overlap: one for prices, one for fundamentals, and ours for what was said on the call. A common configuration is Polygon or Alpha Vantage for market data, FMP or Financial Datasets for statements and ratios, and earningscalls.dev for narrative. Add a broker server only if you genuinely want execution or portfolio context — and if you do, keep it on paper until you trust the workflow.

Connecting ours takes one URL:

https://earningscalls.dev/mcp

Clients that support OAuth, like claude.ai, will walk you through sign-in from there. For config-file clients — Claude Desktop, Claude Code, Cursor — generate a personal connector URL in your dashboard and paste that instead; it carries its own credential, so there is no OAuth round-trip, and each one is revocable per device if a laptop goes missing. The full walkthrough per client is on the MCP setup page. If you would rather call endpoints directly from your own code, the same data is available through the earnings call API, with schemas in the API reference — the MCP versus REST comparison covers when each one is the right choice.

One practical note on running several servers at once: if three of them claim to know something about a ticker, say which source you want. Assistants are reasonable at picking tools, but "search the transcripts for this" beats hoping.

A question only transcripts can answer

Here is the kind of prompt that fails with every other server on this list:

Pull the last four earnings calls for this company. For each one, find what the CFO said about gross margin, quote it, and tell me whether the framing changed. Then find which analyst pushed on margins in the Q&A, and whether they got a straight answer.

Work through what it requires. Listing the call history is one tool call. Speaker segments filtered to executives isolate the CFO's words from the CEO's framing and the operator's boilerplate. Searching within the ticker finds the margin passages without loading four full transcripts into context. The analyst side then needs Q&A segments with attribution intact, because "which analyst asked" is the whole point.

A market-data server can tell you gross margin went from 74.2% to 71.8%. It cannot tell you that the CFO stopped calling the compression "transitory" in the third quarter, or that the same analyst asked about it three calls running and got a longer answer each time. That gap between the number and the explanation is where most of the research work lives. For the same approach under time pressure, see preparing any earnings call in 15 minutes.

Pick your servers on the merits. Brokers handle execution, market-data vendors handle prices, screeners handle screening. For the narrative — what was said, by whom, and how it changed — connect ours.

Related comparisons

Ready to give Claude the transcripts? Connect the server at earningscalls.dev.