Insights
MCP Server for SEC Filing Research
Alphanume Team · August 22, 2026
Use an SEC filing research MCP server to separate filing activity, financing authorization, issuance workflows, defaults, suspensions, and cyber disclosures before joining them into an event cohort.
The main benefit of an MCP server for SEC filing research is not summarizing one filing. It is giving an agent distinct, documented tools for questions that are easy to blur together. A burst of filings is not a shelf registration. A shelf registration is not an offering. A default disclosed in an 8-K is not the same event as an SEC trading suspension. Alphanume keeps these datasets separate and exposes each through the same hosted MCP connection.
The server provides 25 read-only tools backed by the REST API, with source URLs and structured fields where the underlying dataset supplies them. The model chooses and calls tools, but it does not independently validate an event label or guarantee a research result. Require the output to name the dataset, event date, source field, and missing-value policy before combining records.
Choose the filing dataset that matches the claim
Dataset tool | What one row represents | Claim it does not support alone |
|---|---|---|
get_filing_intensity | One ticker-day filing count | That a specific filing is dilutive, adverse, or material |
get_dilution_filings | One S-1 registration event and its lifecycle fields | That registered shares were sold into the market |
get_shelf_registrations | One S-3 or F-3 shelf filing, amendment, or automatic shelf | That registered capacity became issuance |
get_corporate_default_events | One labeled default event from a filing | That the issuer entered bankruptcy |
get_sec_trading_suspensions | One issuer named in a Section 12(k) order | That the security actually resumed quoting later |
get_cyber_incidents | One Item 1.05 filing or amendment | That an unstated impact did not occur |
These distinctions are part of the data contract, not editorial nuance. Shelf capacity_amount is authorization, not issuance. Cyber flags are three-state, so null means the filing did not say. A suspension's resumption_at is the date trading may legally resume, not evidence that a market maker posted a quote. The MCP tool catalog links each tool to its complete field reference.
Build a filing-intensity and shelf cohort
A defensible cross-tool question is whether unusual filing activity coincides with a recent shelf registration. The first dataset is a screening signal; the second supplies the actual financing record. Use a fixed observation window and keep the filing date from each source rather than collapsing both into one vague event date.
Use only Alphanume MCP tools.
Observation window: 2026-07-01 through 2026-07-31.
1. Call get_filing_intensity with filing_count_gte=4 for the window.
2. For the resulting tickers, call get_shelf_registrations for the same window.
3. Keep one table with ticker, intensity date, filing_count, shelf filing date,
form, shelf_type, capacity_amount, is_resale, became_effective,
accession_number, and filing_url.
4. Keep null capacity as null. Do not describe it as zero.
5. Label the output a research cohort, not confirmed issuance.
6. Return the exact tool arguments and any truncation markers.This query can surface a relationship worth testing, but it cannot establish causality. Filing bursts can reflect routine reports, governance changes, transactions, or amendments. The shelf record adds specificity, yet even a large primary shelf can remain unused. A later offering or takedown is a separate event that must be observed rather than inferred.
Store separate clocks and source evidence
SEC-derived datasets carry different clocks. Filing Intensity uses a ticker-day count. Dilution records use the registration filing date and may later acquire effectiveness or withdrawal dates. Shelf rows can update after EFFECT notices or 424B5 takedowns. Cyber records separate discovery, materiality determination, disclosure, and amendment dates. Joining only on ticker can erase the sequence that makes the study meaningful.
Audit column | Why retain it |
|---|---|
observation_date | Defines what the screen could know when it ran |
source_event_date | Preserves the filing, order, or meeting date used by that dataset |
last_updated | Shows when a row may have gained later lifecycle information |
source_url | Allows the structured label to be checked against the primary document |
null_fields | Prevents unstated facts from becoming false or zero |
For lifecycle research, save the row as first observed and refresh it separately. Otherwise a field populated weeks later can leak backward into the original decision date. An updated_since query is useful for collecting changes, but the updated row should not overwrite the historical snapshot used by a prior backtest run.
Access, pagination, and failure modes
- Free access covers a trailing 20-trading-session window delayed by one session; older filing studies require Pro history.
- MCP data tools cap assistant-facing output at 500 rows by default, so narrow the window and inspect truncation markers.
- Several wide endpoints paginate, but their cursor keys differ. Return the cursor object exactly as documented for that dataset.
- Sparse event feeds can legitimately return no rows inside a short window. Check tier and filters before calling it missing coverage.
- A model summary is secondary evidence. Preserve the filing URL and inspect it before treating an extracted amount or label as decisive.
Run the prompt on one month and export two artifacts: the raw tool responses and the joined cohort. Manually check five rows against their filing URLs, then rerun with a lower filing-count threshold to see whether the cohort changes mechanically. Confirm the required history on the access page, and use the corporate-events research hub to choose a later outcome without confusing registration, issuance, default, or suspension.