Best SEC Filings APIs for AI Agents

TL;DR: choose by workflow, not by logo

Who this guide is for

This guide is for engineers building research copilots, event-monitoring agents, earnings workflows, compliance review, or retrieval systems that must answer with primary-source evidence. It assumes you can make HTTPS requests, store normalized JSON, and run a scheduled job or queue.

What an AI agent actually needs from a filing API

“Returns a 10-K” is not a sufficient requirement. A production agent must identify the issuer, distinguish filing dates from reporting periods, preserve amendments, retrieve the primary document and exhibits, and prove which source supported every claim.

• Identity: CIK as the stable SEC key; ticker and legal name as aliases.
• Discovery: filter by form, issuer, filing date, report date, and amendment status.
• Documents: primary HTML or text, filing index, exhibits, inline XBRL, and attachments.
• Evidence: accession number, canonical source URL, retrieval time, content hash, and cited section.
• Operations: documented limits, retries, caching, backfills, and predictable error behavior.

SEC filings APIs compared

Do not choose from a feature matrix alone. Run the same 25–50 filings through every shortlisted provider: recent and historical 10-Ks, 10-Qs, 8-Ks with exhibits, amendments, a renamed issuer, a ticker change, and at least one filing with custom taxonomy facts.

Provider notes: where each option wins

Reference architecture for a source-backed filing agent

Step-by-step: fetch the latest 10-K from EDGAR

The SEC Submissions API is a good baseline because it forces the workflow to preserve first-party identifiers. Use a descriptive User-Agent with organization and contact information, cache responses, and stay below the SEC’s published fair-access ceiling.

curl --compressed \
  -H "User-Agent: ResearchAgent/1.0 contact@example.com" \
  -H "Accept-Encoding: gzip, deflate" \
  "https://data.sec.gov/submissions/CIK0000320193.json"

import hashlib
import requests
from datetime import datetime, timezone

CIK = "0000320193"
HEADERS = {
    "User-Agent": "ResearchAgent/1.0 contact@example.com",
    "Accept-Encoding": "gzip, deflate",
}

submissions_url = f"https://data.sec.gov/submissions/CIK{CIK}.json"
data = requests.get(submissions_url, headers=HEADERS, timeout=30)
data.raise_for_status()
payload = data.json()
recent = payload["filings"]["recent"]

rows = [dict(zip(recent, values)) for values in zip(*recent.values())]
filing = next(row for row in rows if row["form"] == "10-K")

accession = filing["accessionNumber"]
accession_path = accession.replace("-", "")
cik_path = str(int(CIK))
primary = filing["primaryDocument"]
source_url = (
    f"https://www.sec.gov/Archives/edgar/data/"
    f"{cik_path}/{accession_path}/{primary}"
)

document = requests.get(source_url, headers=HEADERS, timeout=30)
document.raise_for_status()

evidence = {
    "cik": CIK,
    "accessionNumber": accession,
    "form": filing["form"],
    "filingDate": filing["filingDate"],
    "reportDate": filing["reportDate"],
    "primaryDocument": primary,
    "sourceUrl": source_url,
    "retrievedAt": datetime.now(timezone.utc).isoformat(),
    "sha256": hashlib.sha256(document.content).hexdigest(),
}

This example intentionally selects only the exact form 10-K. If your workflow should include amendments, model 10-K/A as a separate version and define how it affects earlier conclusions.

Define an evidence contract before prompting the model

Know which filing answers which question

Validation checklist before an agent can answer

• The CIK maps to the intended legal entity, not merely a matching ticker string.
• The accession number has the expected format and resolves to a filing index or primary document.
• The form policy explicitly includes or excludes amendments.
• Filing date, report date, acceptance time, and fiscal period are not conflated.
• Required exhibits were retrieved and linked to the parent filing.
• XBRL facts agree on taxonomy, unit, duration or instant, dimensions, and period.
• Every generated claim has a source URL and exact evidence location.
• The stored hash and retrieval timestamp make the input reproducible.

Security, cost, and latency controls

SEC currently asks automated users to make no more than 10 requests per second in total and may block excessive or unclassified bots. Treat that number as a ceiling, not a target. Use a shared rate limiter, descriptive User-Agent, cache, exponential backoff, jitter, and bulk archives for large backfills.

Common failure modes and fixes

News, transcripts, and prices are supporting layers—not filing APIs

A news API explains market reaction. An earnings-call transcript adds management language. Price and volume data measure response. None of them replaces the filed document, accession number, amendment history, or SEC source URL. Retrieve the filing first, then join these datasets using issuer identity and event time.

QVeris implementation pattern: Discover → Inspect → Call

In an agent system, QVeris can sit above provider-specific endpoints as a discovery and routing layer. The agent discovers candidate SEC tools, inspects the input schema and provider metadata, then calls the selected tool with bounded arguments. The returned evidence object should still name the underlying provider and retain the original SEC source.

Frequently asked questions

Official sources and next steps

Provider features and access rules can change. Verify current documentation, entitlements, and terms before production use.