Tested head to head on the same companies and quarters, the three routes agree on the numbers almost everywhere they overlap. What separates them is coverage, traceability and structure, and on those they are not close.
Use a library as the first call - edgartools for most of this. One call like Company("AAPL").get_financials() against a dozen-plus hand-built API requests is not a close contest, and the same holds for statements, filings, notes, holdings, insider forms and text search. It wins on effort and, more importantly, on the number of ways you can quietly get something wrong. It is open source (MIT, by Dwight Gunning) and ships an MCP server plus agent skills, so an AI agent can drive it directly.
Keep the raw SEC XBRL API for three specific jobs. Cross-sectional screening across thousands of filers in one call, which has no library equivalent. Dependency-free one-off lookups. And second-path verification - the value there is precisely that it does not share the library's code or its assumptions.
Treat yfinance as a quick look, not a source. Its numbers reconcile, but it silently drops older quarters from a rolling window, normalises period-end dates to month-end (convenient, slightly false), flattens the statement hierarchy, and drops the audit trail entirely - no accession number, no form type, so you cannot trace a figure back to the filing. That last one is disqualifying if you intend to publish.
The principle worth keeping: pull from two independent paths whenever the number has to be right. In this exercise exactly one figure disagreed across sources, and it turned out to be an error in handling a restatement - a mistake no amount of internal checking within one source would have caught. Two-path reconciliation is what surfaced it.
The how-to for the library route: https://www.tigzig.com/agents-faq/how-to-download-sec-edgar-filings-in-python. Full write-up with every script and dataset: https://www.tigzig.com/story/sec-xbrl-three-ways/.
← All Agents FAQ