We Built EM Data MCP Because EM Credit Workflows Are Still Manual

Most EM credit analysts spend more time hunting for numbers than judging them. That is the problem we set out to close not with better scraping, but with structured, source-traceable data extracted directly from filings.

Open a major EM issuer's annual report. The debt breakdown is on page 187. Revenue by segment is around page 90. The EBITDA reconciliation is in the investor presentation, not the annual report. The maturity schedule is buried in the MDA, from three quarters back, in a table format that does not match the latest filing.

Now do that for five comparable issuers across three countries, under different accounting standards, with some filings switching reporting currencies mid-series without explanation.

That is the daily reality of EM high-yield credit analysis. And it has not changed in twenty years.

"In credit analysis, a number without a source is just a guess with formatting."

Why EM data is uniquely broken

Developed market credit analysts have Bloomberg, CapIQ, and a dozen other providers with clean, standardized data. The filings follow predictable formats. Aggregators pick up numbers quickly.

Emerging markets are structurally different. EM issuers often report under local GAAP then restate under IFRS. Some publish in English, many do not. Disclosure depth varies wildly — one issuer gives you a full tranche-level debt schedule with coupon rates; the next buries its borrowings in a two-line note.

The providers that cover DM either skip EM names entirely or cover them with significant lag and missing fields. If you have pulled an EM issuer on a terminal and found half the fields blank, you know exactly what this looks like.

The default response: Analysts go directly to filings. Open the PDF, find the table, manually build the model in Excel. It works. It takes hours. It does not scale. When a client calls asking for a peer comparison across six names, you are looking at a full day of work just to get the numbers into a usable format.

Where LLMs fall short

The obvious next step — and what many teams have started doing — is attaching annual reports to Claude or ChatGPT and asking for the extraction. For simple, one-off tasks, this works reasonably well.

The cracks show up fast. Try attaching four years of annual reports for a single issuer: easily 800 to 1,000 pages. Most models hit context window limits before you get through all of it. You end up with FY2024 numbers but missing FY2022. Or the model pulls revenue from one filing and EBITDA from another, silently mixing restated and originally reported figures.

And even when extraction works, there is a deeper problem: you get a number, but you cannot answer where it came from, which filing, which page, whether it was restated. You cannot use that in an IC memo or a client deliverable. Not responsibly.

‍

What we built and why

EM Data MCP is a structured database of financial data extracted directly from emerging market company filings. Not web-scraped. Not pulled from third-party aggregators. Extracted from the actual PDFs, page by page, table by table.

Coverage currently spans around 100 EM issuers across Latin America, the Middle East, Africa, CEE, and Asia — expanding to 500 to 700. For each issuer: income statements, balance sheets, cash flows, revenue and cost breakdowns by segment, EBITDA reconciliations, debt structures by seniority and tranche with coupon rates and currency detail, maturity schedules, and forward guidance. Typically four to five years of annual history plus the last eight quarters.

Every data point carries a source reference  the filing name, the page number, and whether the figure was restated. No ambiguity. No derived estimates, no modeled numbers, no proprietary adjustments you cannot trace. If a company reports adjusted EBITDA, we extract it. If they do not, we do not compute it.

The MCP server connects to Claude or any LLM that supports the protocol. You ask a question in natural language. The model queries the database. You get back a structured, sourced answer — no PDFs to upload, no context window limits, no stitching together partial outputs.

How analysts actually use it

1. Getting up to speed on a new name. A PM flags a new issuer. Instead of half a day building a model just to get oriented, you query the MCP and have the income statement, EBITDA reconciliation, and debt structure across multiple years all sourced in seconds.

‍2. Peer comparison tables. The most common request in EM credit. Building manually means opening 15 to 20 filings across multiple issuers, normalizing for currency, assembling everything. With the MCP, you ask Claude to compare capital structures or leverage metrics across issuers and get a clean table back. Same source rigor.

‍3. New issue analysis. 24 hours, sometimes less. You need historical financials, current capital structure, where the new bond fits in the stack, and comps at similar ratings. The MCP gives you the financial profile and complete debt structure immediately — you focus on whether the deal is priced fairly, not on pulling page 214 of an annual report.

‍4. IC memos and client reporting. Every number needs to be defensible. MCP output gives you source filing and page number by default. You can reference them directly in the memo or verify on challenge. It removes the low-grade anxiety of "did I pull this from the right place?"

‍5. Portfolio monitoring. Every quarter, new filings, new numbers to update. Instead of reopening 20 models and manually updating each one, you query the MCP for the current picture. For a 30-name portfolio, that is the difference between a week-long update cycle and an afternoon.

See it in action

‍

Why traceability is non-negotiable in EM

In DM credit, cross-checking is easy. Multiple providers carry the data, filings are standardized, consensus is well-covered. If a number looks off, you verify it in seconds.

In EM, that safety net does not exist. When you extract a number from an EM filing, you need to know exactly where it came from because there may be no other way to verify it. The moment you blend sourced data with derived estimates, you lose the ability to fully trust the output. We decided early that purity of sourcing was more important than covering every possible metric.

‍

That is not a limitation. It is the point.
‍