Few public-market data sources pack as much signal into as little structure as earnings calls. Four times a year, management teams deliver prepared remarks and then take unscripted questions from analysts - and what gets said, and how, moves prices and reshapes models well before it surfaces in reported numbers. None of this is hidden. The hard part is turning a public audio-and-text artifact into something a machine can query at scale, across every name you follow.
This guide covers the full landscape as it stands in 2026: the two programmatic access models - the classic REST API and the newer Model Context Protocol (MCP) connector - what serious coverage actually requires, how to set up each path, and the workflows they enable. It is written for quants building screeners, developers wiring transcripts into retrieval-augmented generation, and analysts who simply want Claude to read the latest call for them.
Earnings call transcripts, and why an API matters
An earnings call is the quarterly (occasionally semi-annual) conference call where a public company discusses results. Its transcript is the verbatim record: executive prepared remarks, operator housekeeping, and the analyst Q&A - the section where most of the signal hides, precisely because nobody scripted it.
Reading one transcript is trivial. The problem starts at scale. Comparing margin language across 40 companies, counting "macro headwinds" mentions in a sector over eight quarters, or piping every fresh call into an agent each morning - none of that works with copy-paste. What it requires:
- Structure - transcripts decomposed into speaker segments with names and roles, not raw HTML.
- Breadth - coverage well beyond a handful of US mega-caps.
- History - enough back-catalog to measure change, not just state.
- Freshness - a call that ended twenty minutes ago should already be queryable.
An API or MCP connector is what converts a pile of documents into a dataset you can filter, screen, and reason over. Once transcripts are structured and queryable, entire categories of analysis open up that are simply not feasible manually - cross-company language tracking, systematic tone comparison, and automated monitoring across thousands of names at once.
REST API vs MCP connector: two access models
The two ways to consume transcript data programmatically serve genuinely different jobs.
The REST API is the traditional integration: your code sends HTTP requests with an API key and gets JSON back. Deterministic, scriptable, and right for backends, cron jobs, dashboards, and pipelines - you decide exactly what is fetched and when.
The MCP connector is the newer model. MCP is an open standard that lets AI assistants invoke external tools directly. Rather than writing request code, you paste one connector URL into your client, and the model gains tools it calls on your behalf - fetch NVDA's latest call, search transcripts for supply-chain commentary, list this quarter's tech reporters. Your natural-language question determines which tools fire.
| Dimension | REST API | MCP connector |
|---|---|---|
| Caller | Your code | An AI assistant (Claude, Cursor, ...) |
| Interface | HTTP + JSON | Model-invoked tools |
| Auth | X-API-Key header |
Token inside the connector URL |
| Suited to | Pipelines, backends, batch jobs | Interactive research, agents, ad-hoc Q&A |
| Determinism | You control every request | Model selects the tools |
| Setup | Write and maintain client code | Paste one URL |
| Learning curve | Standard REST | None - plain language |
They are complementary, not rivals. A common arrangement: REST powers scheduled ingestion while the MCP connector lets analysts explore the same data conversationally, all on one account.
What good coverage actually means
Transcript providers differ enormously, and four dimensions determine the value.
Breadth. A source capped at the S&P 500 fails the moment a European industrial or Asian semiconductor supplier matters to you. The earningsapi.io dataset covers 12,728 companies across all 11 GICS sectors, spanning dozens of countries and exchanges worldwide - global coverage, not just US large-caps.
History. One quarter is a snapshot; trends need years. Coverage runs from 2020 to the present, totaling 250,945 earnings calls.
Speaker structure. This separates a dataset from a wall of text. Every transcript is decomposed into speaker segments - 11.9M of them - each classified as executive, analyst, operator, attendee, or shareholder, with named speakers on the vast majority. Filtering to "everything the CFO said" or "all analyst questions" is what makes screening and RAG feasible at all.
Freshness. New calls become queryable within minutes of ending, and cursor-based polling lets you fetch only what appeared since your last check.
Evaluate any provider against all four. Depth without breadth confines you to one market; breadth without structure leaves you parsing text by hand; freshness without history rules out trend analysis.
Working with the REST API
The API is the foundation for anything automated.
- Base URL:
https://earningsapi.io/api/v1 - Auth: your key in the
X-API-Keyheader on every request. - Spec: a machine-readable OpenAPI definition at
https://earningsapi.io/openapi.json- generate a typed client in any language.
The core surface: an endpoint to list and filter calls (ticker, date, sector, and more), a latest-calls endpoint built for polling, full-text search across transcripts and segments, and a transcript-by-ID endpoint that returns the structured speaker segments.
A minimal request:
curl -H "X-API-Key: your_key_here" \
"https://earningsapi.io/api/v1/earnings/latest"
Two capabilities matter once you scale. Cursor-based polling keeps ingestion jobs cheap and idempotent by returning only calls added since your last request. And batch ticker support on aggregate endpoints lets you pull an entire watchlist in one call instead of iterating request by request. Full parameters and examples live in the docs; everything is versioned under /api/v1, so integrations stay stable.
Working with the MCP connector
Inside an AI assistant, the connector is dramatically faster than writing client code. The MCP server runs at mcp.earningsapi.io, and your personal connector URL - generated from the dashboard - looks like:
https://earningsapi.io/u/mct_xxx/mcp
The access token is embedded in the URL, so treat it as a secret. Then register it in your client:
- Claude Desktop - Settings → Connectors, add the URL.
- claude.ai (web) - add it as a custom connector in account settings.
- Claude Code -
claude mcp add --transport http earningsapi https://earningsapi.io/u/mct_xxx/mcp - Cursor - add an entry named
earningsapito your MCP configuration.
Connected, the model gains a toolset mirroring the REST surface: latest call by ticker, call listings per company, transcript search, speaker segments, sector and industry browsing, upcoming earnings, and more. You ask in plain language - "summarize Apple's most recent call and flag any guidance changes" - and the model chains the right tools. The /mcp page has the per-client walkthrough and current tool list.
What people build with it
One dataset, very different workflows.
Fundamental research. Summarize and cross-reference calls without leaving your assistant - guidance deltas, every mention of a product line, tone shifts across quarters. The connector turns questions that used to mean an afternoon of reading into a two-minute exchange.
Market-wide screening. Instead of reading one call, sweep thousands: every company in a sector that raised guidance, every mention of a specific supplier, every management team that suddenly stopped talking about buybacks. Speaker structure plus batch ticker queries make this practical rather than theoretical.
RAG pipelines. Load transcripts into a vector store for grounded Q&A over years of calls. Because segments arrive pre-chunked by speaker and tagged with ticker, sector, and date, retrieval quality starts high and citations come for free.
Autonomous agents. An agent pulls a company's latest call, diffs it against prior quarters, and writes up what changed - end to end, unattended. The self-describing MCP toolset means the agent needs no hand-built integration layer to do it.
Fresh-call alerts. Minute-level freshness plus cursor polling supports an agent that watches for new calls and pings you when a followed company reports or a keyword appears - useful during earnings season, when dozens of relevant names can report in a single week.
Evaluating a provider
A checklist worth applying to any transcript source:
- Global breadth - does coverage reach the exchanges and countries you actually track?
- History - enough back-catalog for time-series work?
- Role-classified speaker segments - filterable by role and named speaker, or undifferentiated text?
- Freshness and incremental fetch - how fast do calls land, and can you pull only the delta?
- Both access models - REST and MCP, so no workflow lock-in?
- A published OpenAPI spec - typed clients in minutes, not days.
- Batch endpoints - whole watchlists per request, not thousands of tiny calls.
EarningsAPI was built against exactly this list: 2020-to-present history, 250,945 calls, 11.9M role-classified segments across 12,728 companies, minute-level freshness, cursor polling, batch ticker queries, an OpenAPI spec, and both a REST API and an MCP connector on one account.
Plans and first steps
Access is subscription-based with quarterly billing:
| Plan | Price | For |
|---|---|---|
| Basic | $105 / quarter | Individual projects and evaluation |
| Pro | $145 / quarter | Heavier research and production use |
| Ultra | $515 / quarter | High-volume pipelines and agents |
Exact request limits per plan are listed on the pricing page. Getting going takes four steps:
- Pick a plan on the pricing page.
- In the dashboard, copy your API key (REST) or generate an MCP connector URL.
- Follow the docs for the API, or the /mcp guide for the connector.
- Make your first request or ask your first question.
FAQ
REST API or MCP connector - which should I start with? REST for backends, pipelines, and scheduled jobs that need deterministic control. MCP when an AI assistant should research and reason over transcripts conversationally. Both run on the same account, so many teams simply use both.
How current is the data? New calls are typically queryable within minutes of ending. Cursor-based polling returns only what is new since your last request.
What does coverage include? 250,945 earnings calls from 12,728 companies across all 11 GICS sectors, with history back to 2020 and 11.9M speaker segments classified by role.
Do MCP and REST see different data? No. Both interfaces read the identical corpus under the same account and key - only the access pattern differs.
Wrapping up
The distance from "I need earnings call data" to a working integration is short: choose a plan, generate a key or connector from the dashboard, and within minutes five-plus years of speaker-structured transcripts are available to your code or your AI assistant.