It's the second week of earnings season. You have eleven names to cover, four reporting tomorrow morning, and a portfolio manager who wants a one-line read on each before the open. The transcripts are 45 minutes long. The prepared remarks bury the number you actually care about somewhere between a safe-harbor statement and a thank-you to the operator. The Q&A is where the real information lives, but that's another 30 minutes of scrolling.

This is the analyst's structural problem: the information you need is public, verbatim, and sitting in a transcript, yet extracting it is slow, manual, and does not scale across a coverage list. During peak weeks, the bottleneck is not your judgment. It's reading speed.

AI earnings analysis changes the economics of that bottleneck. When Claude can query a transcript database directly through MCP, you stop reading calls end to end and start interrogating them. You ask a specific question and get a cited answer, with the verbatim numbers, in seconds. This article walks through exactly how that works, why the citations matter more than the speed, and how to set it up in about five minutes.

The old workflow

Here is what covering a single earnings call looks like without tooling:

  1. Find the transcript. Search a provider, wait for it to post, hope your data vendor covers the name and the exchange. For non-US listings this alone can eat an afternoon.
  2. Read it. 45 minutes if you read fast and don't get interrupted. You skim prepared remarks, then slow down for Q&A because that's where guidance gets pressure-tested.
  3. Take notes. You copy the revenue number, the margin comment, the guidance change, the one analyst question that got a defensive answer. You paste them into your model or your notes doc.
  4. Repeat. Now do it ten more times this week. And do it again next quarter when you need to remember what management said last time.

The math is brutal. Ten names at roughly an hour each of find-read-note is a full workday gone to mechanical extraction before you've formed a single opinion. Cross-quarter comparison — "did they walk back the FY guide they gave in Q1?" — means reopening old transcripts and re-reading them, because your notes never capture everything. And a sector-wide read ("who mentioned pricing pressure this quarter?") is effectively impossible by hand across a whole coverage universe.

The MCP workflow

MCP (Model Context Protocol) lets Claude call external tools. Connect Claude to earningscalls.dev's MCP server and Claude gains direct, structured access to the transcript corpus: 246,000+ earnings call transcripts covering 12,000+ companies across 70 countries, 170+ exchanges, and all 11 GICS sectors, from 2020 to the present. That's 11M+ speaker segments Claude can search, filter, and quote.

The workflow inverts. Instead of reading a call to find an answer, you ask the answer directly:

Pull Nvidia's latest earnings call. What did management say about data center revenue growth and gross margin guidance? Quote the exact figures with speaker attribution.

Claude runs a couple of tool calls in the background — find the company, get the latest call, pull the relevant speaker segments — and comes back with the verbatim numbers, who said them, and where in the call. No scrolling. A typical research session is 3-5 tool calls, and it finishes in the time it used to take you to locate the transcript.

The 10x is not marketing. If find-read-note was an hour and a targeted query is a few minutes, the per-call time collapses by roughly an order of magnitude — and the savings compound the moment you need to compare across quarters or sweep across a sector, because those tasks go from impractical to a single prompt.

A day in earnings season with Claude + MCP

Here's a concrete run through a reporting day.

Pre-call prep (morning, before the open). You have four names reporting today. Before each call, you want the setup: what did management guide to last quarter, and what are the known sore spots? One prompt per name:

For [ticker], summarize the FY guidance management gave on the last two earnings calls and flag anything they hedged or revised. Quote the guidance language verbatim.

Now you walk into each call knowing exactly what to listen for and what a beat or miss against management's own words would look like.

Post-call debrief (mid-morning). The call just ended. The transcript posts. Instead of reading it, you ask:

Get [ticker]'s call that just posted. Compare what management said today about margins and demand to what they said last quarter. What changed in tone or numbers?

You get a debrief with the verbatim before-and-after in the time it takes to refill your coffee.

Cross-quarter compare (midday). A PM asks whether a name has quietly walked back a target it set earlier in the year. This is the query that was painful by hand:

Across [ticker]'s last four quarterly calls, track how management talked about their operating margin target. Show the exact language from each call in date order.

Claude pulls each call, extracts the relevant segments, and lays them side by side with dates and speakers. You see the drift — or confirm there isn't any — in one pass.

Sector sweep (afternoon). Your PM wants to know who's flagging pricing pressure across your semiconductor coverage this quarter. Instead of ten transcripts, one search across the corpus surfaces the companies and the exact quotes where executives raised it, so you can rank the names by how loudly management is worrying out loud.

Four tasks that used to fill a day. Done between other work, each backed by verbatim source text.

Why citations matter

Speed is the headline, but for an analyst the real product is trust in the answer. This is where AI earnings analysis usually breaks down: a chatbot summarizing a news article about a call is summarizing someone else's interpretation, one step removed from what management actually said — and a step where numbers get rounded, context gets dropped, and errors get introduced.

With the MCP connection, Claude reads the actual transcript text. The primary source. When it tells you data center revenue grew a specific percentage, that figure comes from the words the CFO spoke on the call, and Claude can quote it verbatim with speaker attribution — executive, analyst, or operator.

That matters for three reasons an analyst will recognize immediately:

Citations turn Claude from a fast-but-unaccountable summarizer into a research assistant whose work you can check. That's the difference between a toy and a tool you put in your workflow.

Setup in 5 minutes

There's no data pipeline to build. You connect Claude to the MCP server once.

  1. Go to your dashboard → Connectors → Generate. You get a personal MCP URL that looks like https://earningscalls.dev/u/mct_xxxx/mcp.
  2. Claude Desktop or claude.ai: Settings → Connectors → Add custom connector → paste the URL.
  3. Claude Code: run claude mcp add --transport http earningscalls "<url>".

That's it. Claude now has the transcript tools. Start asking. If you want a deeper walkthrough with screenshots, see How to connect earnings call transcripts to Claude with MCP, and for ready-to-paste queries, 10 Claude prompts for analyzing earnings calls via MCP. The full tool reference lives in the docs. If your work leans toward screening rather than single names, Screen earnings calls across the market with MCP tools covers the sector-sweep patterns in depth.

What it costs vs the time saved

Reading transcripts on the website is free — start there to see the coverage. The MCP connection, which is what makes Claude fast, runs on usage-based plans:

Each tool call is one request, and a typical research session is 3-5 tool calls. On Pro, that's roughly a thousand research sessions a month before you hit the ceiling. On Ultra, you can run heavy sector sweeps and cross-quarter comparisons across a full coverage list every single day without thinking about it.

Now weigh that against the alternative. If AI earnings analysis saves you even a few hours during a single reporting week — and if it collapses a day of find-read-note into a morning, it saves far more than that — the plan pays for itself many times over in the first week of the quarter. For a research professional whose time is the scarce resource, this is not a close call.

Read earnings calls faster, starting today

The transcripts are public. The numbers are verbatim. The only thing standing between you and a cited answer in seconds is the connection. Generate your MCP URL from the dashboard, paste it into Claude, and turn your next reporting day from a reading marathon into a series of questions.

See the plans on pricing and start covering your list at the speed of a prompt.