The default way to get an AI assistant to reason about an earnings call looks like this: locate the transcript, copy the relevant chunk, paste it into the chat, ask your question. Need a second quarter for comparison? Repeat. Different company tomorrow? Start from zero.
It's tedious, and it doesn't scale. Context windows cap how much you can paste, and there is no pasting your way to "search every semiconductor call from last quarter." The model is capable; it just can't see the data.
An earnings call MCP server removes you from that loop. You connect your assistant to a transcript database once, and afterwards it retrieves whatever it needs on its own, mid-conversation. This guide covers what that means, how the hosted EarningsAPI server works, and how to wire it up.
The protocol, briefly
MCP is the Model Context Protocol - an open standard for connecting AI assistants to external tools and data in a uniform way.
The useful mental model is a standardized plug. An MCP server publishes a set of callable tools (search, fetch, list). An MCP client is software you already run: Claude Desktop, claude.ai, Claude Code, Cursor. Any conforming client can connect to any conforming server, discover its tools automatically, and invoke them without custom integration work on your side.
Applied to earnings calls
An earnings call MCP server is simply an MCP server whose entire toolset revolves around transcript data: searching calls, fetching them, slicing them by speaker, checking the earnings calendar.
The shift is who does the retrieval. You stop being the courier between database and model. You ask a question in plain language; the assistant selects a tool, pulls the exact transcript or segment it needs, and answers with the source text in front of it - instead of guessing from training data that ended months ago.
The MCP setup guide covers the protocol background in more depth.
How the EarningsAPI server works
EarningsAPI operates a hosted remote MCP server at mcp.earningsapi.io. Nothing runs on your machine - no local process, no package install. You add a URL to your client, and search, transcripts, and metadata are served from our side.
The dataset behind it:
- 250,945 earnings call transcripts, 2020 through today
- 12,728 companies across global exchanges and every GICS sector
- 11.9M speaker segments, each tagged by role
That role tagging is the quiet workhorse. Every segment is labeled executive, analyst, operator, attendee, or shareholder, which is what turns "what did the CFO say about margins" from a fuzzy request into a precise filtered query.
The toolset
The server publishes around a dozen tools. You never invoke them manually - the assistant picks based on your question. Functionally, they cover:
- Full-text search across the entire corpus
- Fetching a complete transcript for any call
- Pulling speaker segments, filtered by role
- Listing a ticker's call history, or just its latest call
- Upcoming earnings dates
- Company lookup by name
- Browsing sectors and industries
- Dataset-wide statistics
Because each tool is small and structured, the assistant chains them freely: resolve a company, locate its newest call, extract the executive remarks - all triggered by one question from you.
Questions that just work
Once connected, you talk to your assistant the way you already do:
"What did NVIDIA's CFO say about data center growth on the latest call?"
The assistant resolves the ticker, finds the most recent call, filters to executive segments, and quotes the actual transcript back to you. Zero copy-paste.
More examples from real usage:
- "Summarize the risks management raised across Microsoft's last three calls."
- "Which analysts pressed Tesla on margins last quarter, and how?"
- "Contrast how two rivals framed AI spend this quarter."
- "When does Apple report next?"
- "Find every 'supply chain' mention in recent semiconductor calls."
The division of labor is simple: the model fetches, you think.
Who gets value from this
Anyone who spends real hours on earnings calls:
- Independent investors who want sourced answers rather than PDF skimming.
- Analysts comparing management language across peers and quarters.
- Developers and quants wiring transcript data into agents and pipelines.
- Finance and strategy teams running competitive monitoring.
If Claude or Cursor is already open on your desk, this makes your existing assistant fluent in earnings calls with no code written.
Setup
The hosted connector is the fast path. Your access token is embedded in the URL itself, so there's no API-key header to configure anywhere.
1. Generate a connector URL
Sign in and open Dashboard → Connectors → Generate. Your personal URL looks like:
https://earningsapi.io/u/mct_xxx/mcp
Treat it like a password - the URL is the credential. You can revoke and regenerate it from the dashboard at any time.
2. Add it to your client
Claude Desktop or claude.ai: Settings → Connectors → Add custom connector, paste the URL.
Claude Code: one command:
claude mcp add --transport http earningsapi "https://earningsapi.io/u/mct_.../mcp"
Cursor and other MCP clients: add a custom HTTP MCP server and point it at the same URL.
3. Ask something
Done. Pose an earnings question and watch the tool calls happen on their own. The step-by-step version with client specifics lives on the MCP page.
MCP or the REST API?
Rule of thumb: MCP when a model is asking, REST when your code is asking.
Chatting with an assistant that reads calls? Use the connector. Building a pipeline, dashboard, or backend job? The REST API serves the identical corpus with full programmatic control. Plenty of users run both against the same plan.
Pricing
API and MCP access share one quota and are billed quarterly:
- Basic - $105 per quarter
- Pro - $145 per quarter
- Ultra - $515 per quarter
Every tool call the assistant makes counts as one API request, so a question that chains three lookups consumes three requests. Full quota details are on the pricing page.
Bottom line
An earnings call MCP server flips your assistant from something you feed transcripts into something that fetches them itself. Connect once, then ask plain-language questions and get answers grounded in what executives and analysts actually said - across 250,000+ calls and 12,728 companies.
To try it, generate a connector URL and follow the setup guide; you'll be asking questions within a couple of minutes.