How to get SEC filings as data

The short way

Every filing a company has made comes back as one JSON document, with no key and no account — but the request has to introduce itself:

import requests

headers = {"User-Agent": "Your Name your.email@example.com"}

cik = "0000320193"  # ten digits, leading zeros included
url = f"https://data.sec.gov/submissions/CIK{cik}.json"

filings = requests.get(url, headers=headers).json()
recent = filings["filings"]["recent"]

# parallel arrays, newest first
for form, date, accession in zip(recent["form"], recent["filingDate"], recent["accessionNumber"]):
    print(date, form, accession)

Two things in there are doing more work than they look. The User-Agent header is not decoration: the SEC's own sample is Sample Company Name AdminContact@<sample company domain>.com, alongside Accept-Encoding: gzip, deflate, and it asks you to declare it on every request. And the CIK is zero-padded to ten digits, because EDGAR is organised by CIK rather than by ticker — the mapping file is a separate download keyed by an index position, with cik_str, ticker and title in each entry, and joining it is your job.

If it is the reported numbers you want rather than the filing index, the XBRL endpoints return them already tagged: companyconcept for one concept at one company, companyfacts for everything one company has tagged, and frames for one concept across every filer in a period.

What the options are

The source, and nothing between you and it. SEC EDGAR is the filing system itself, and four surfaces matter to somebody building rather than browsing: the submissions endpoint per company, the three XBRL endpoints, full-text search from 2001 onward, and the quarterly flat files. The SEC gives a typical processing delay of under a second for submissions and under a minute for the XBRL APIs, which makes this the fastest route in the catalogue as well as the cheapest one. What it does not give you is a parser.

Bulk, when you want the corpus rather than a company. The same product publishes nightly archives of the submissions and company-facts datasets, refreshed at about 3:00 a.m. Eastern. Anything that wants everything belongs here rather than in a loop over endpoints, and the arithmetic is the reason: ten requests a second against tens of thousands of registrants is measured in days.

A commercial layer over the same filings. sec-api.io sells the parsing: a query API over 20 million filings from 1993 onward with twenty-odd filterable properties, an extractor that returns a named item out of a 10-K rather than a document you then parse, an XBRL-to-JSON converter, and a websocket that pushes new filings. From $55 a month, $49 annually, on a single-key personal licence. Read the tier table before the feature list — full-text search, Form ADV, 13D/G, N-PORT, N-CEN and N-PX are all marked unavailable on the cheap plan, and redistribution exists only on an unpriced Enterprise tier.

An assistant doing the reading. SEC EDGAR MCP is an AGPL-3.0 server over these same free endpoints, exposing twenty-one tools: ticker-to-CIK lookup, filings by form and date, 10-K and 10-Q section extraction, an 8-K item reader, XBRL statements, and Forms 3, 4 and 5 parsed into transactions. It requires SEC_EDGAR_USER_AGENT with a real name and email, and defaults to the SEC's ten-per-second ceiling. The licence is the commercially material fact: network-serve a modified copy and its users are owed that source, and the alternative is a commercial licence with no published price. The rest of that shelf is in MCP servers and AI agents.

One form, done properly. If the question is really 13F, WhaleWisdom has it parsed back to the quarter ending 31 March 2001 behind a signed REST API, with N-PORT constituents since 2019 alongside. A registered free account reaches the last eight quarters; Standard at $300 a year opens the history and meters it at 50 funds and 50 stocks per 90 days. 13F.info is free and has no API at all, but it is MIT-licensed and the ingestion code is public, which is a different way to end up with the same table. OpenInsider does the Form 4 half free with a 40-field screener and, again, no API. The rest are in insider, 13F and congressional trade trackers.

The numbers inside the filing rather than the filing. Daloopa extracts segment splits, operating KPIs, guidance and GAAP-to-non-GAAP bridges out of filings, press releases and investor presentations, each figure hyperlinked to the page it came from. That is the material XBRL never tagged, and it is quote-only — the free account is one data sheet download inside a two-week window.

Where this breaks

Ten requests a second, and a header that names you. The SEC publishes a maximum access rate of ten requests per second and says it is carefully monitored to preserve equitable access. It reserves the right to limit request rates, and says that any request identified as part of a botnet or an automated tool outside the acceptable policy will be managed to ensure fair access for all users. There is no key to lose, which means the only thing that can be withdrawn is your address's access. A single-threaded backfill that politely sleeps is the correct shape here, and anything wider should be reading the nightly archives instead.

The XBRL endpoints are financial facts, not filings. They cover the 10-Q, 10-K, 8-K, 20-F, 40-F and 6-K series and their variants, which means a great deal of what people actually want is absent by design: Forms 3, 4 and 5, 13F holdings, 13D and 13G stakes, and — inside the forms that are covered — the risk factors, the MD&A and every other block of prose. Those are the document, and getting them means fetching and parsing it, which is precisely what the paid layers and the MCP server sell.

The submissions endpoint is not the whole history. Its compact arrays hold at least one year's filings or the 1,000 most recent, whichever is more. If the entity has filed more than that, a files array points at additional JSON documents with the date range each one covers. A script that reads recent and stops is a script that silently truncates the oldest filer in your universe while working perfectly on the newest.

Amended and restated are two different problems. The first is easy to see and easy to miss: the form array carries the suffix, so 8-K/A, 3/A and 144/A sit in the index next to the originals, and neither retracts the other. Filtering on form == "10-K" drops every amendment; filtering on startswith("10-K") double-counts the year. The second problem is quieter. The frames endpoint aggregates one fact for each reporting entity that is last filed and most closely fits the period requested, and the SEC warns that users should be mindful of different reporting start and end dates for facts inside a frame — so a frame is a mix of vintages, not a snapshot of what was known on a date. The Financial Statement Data Sets are the other extreme: presented without change from the as-filed reports, with no standardisation and no restatement applied. Whether you want as-first-reported or as-currently-restated is a decision you have to make, and neither surface makes it for you.

Full-text search starts in 2001, and it is a search. It covers all EDGAR filings submitted electronically since 2001, including attachments and exhibits, which is genuinely more than most people expect. What it is not is a query language: natural-language search is unsupported, and the wildcard cannot lead or sit inside a word, or appear in an exact phrase or a boolean search at all. Anything filed before 2001 is in the archives as a document you can fetch by CIK, form and date, and not as text you can search.

The flat files land a quarter behind. The convenience is real — Forms 3, 4 and 5 flattened from XML, 13F holdings as tables — and so is the lag. Insider Transactions run from January 2006 and Form 13F from July 2013, both updated quarterly, with the 13F sets run for the prior three months following the end of February, May, August, November; the Financial Statement Data Sets start in January 2009 and are quarterly too. The SEC attaches the same disclaimer to all of them: the data is derived from what individual filers provided, accuracy is not guaranteed, and the sets are not a substitute for the filings. So the choice is a parser plus sub-second latency, or no parser and up to a quarter of delay. There is no third option that is both.

Nothing is normalised across filers. A 13F is what the manager typed. Names, CUSIPs and share classes disagree between filings of the same security, and a company's own XBRL tag choices drift across years. Reconciling that is not a bug in EDGAR — it is the work every paid vendor on this page is actually charging for.

If you outgrow this

When the delay is the problem, the next purchase is a push rather than a poll. sec-api.io streams new filings over a websocket, and its own pricing page puts that at under 300 milliseconds after a filing becomes public; Daloopa offers webhooks for new and restated data, which is the same idea aimed at the numbers.

When you want statements rather than documents, stop parsing. Financial Modeling Prep, Intrinio and Fiscal.ai sell standardised financials by API for a published monthly figure — fine for an income statement, and useless for the metric that only ever appeared on slide 16. The rest of that shelf is in fundamental data and research platforms.

When the question is timing rather than content, the answer is not a vendor. When a trade shows up in the public record sets out the deadline behind each form — two business days for a Form 4, 45 days after quarter end for a 13F — because no API can hand you a filing that has not been made yet.

When a person is going to read it, BamSEC is the opposite trade: no API at all, $69 a month billed annually, and filings relabelled with table downloads for a human doing the reading. Quartr and the Quartr API sit beside it for earnings calls, transcripts and slide decks, which are the company material EDGAR never held.