eCFR: Search US Federal Regulations, Parts & Amendments

Search the Code of Federal Regulations by phrase, title, part or agency and get sections as rows: citation, headings, matched excerpt, effective window and change types. Further modes list the parts and sections inside a title, amendment dates per section, and the title and agency dictionaries.

Every row comes from the government's own eCFR JSON API — no key, no proxy, no browser, no login. You get the citation (21 CFR 820.10), the part and section headings, the matched passage, the dates a version started and ends, whether it is in force, and the link to the official page; the amendment modes add the amendment and issue dates of every recorded change. Built for compliance and quality teams, regulatory lawyers, policy analysts and agents that need a citation with a link rather than a web page. It returns metadata, structure and excerpts — not the full text of a section; see Limits & FAQ for why.

Need it tailored to your workflow? Missing a field, a filter, another source or a different output format? Open an issue and describe what you need — we read every request and adapt the actor.

Use cases

• Watch the parts your product is regulated under. Run search with lookbackDays: 7 and onlyNew: true on a schedule and the dataset holds only the sections that actually moved this week — with the date the new text took effect.
• Prove when a rule last changed, for an audit. versions on a title and part gives one row per recorded amendment with amendmentDate, issueDate and whether the change was substantive, so a finding gets a dated citation.
• Build a table of contents for a RAG index. structure on a title or part writes every subpart and section with its heading, the range it covers and sizeBytes, so you know how long each node is before you fetch anything.
• Scope a topic before you read. counts turns a phrase into a map of the code — how many matching sections sit in each title, chapter and part — for one request instead of thousands of paid rows.
• Give an agent a citation tool. One run-sync-get-dataset-items call answers "which section says this" with the citation, the heading and a link a human can open.
• Keep a dated copy of the code for a contract. asOfDate reads the code as it stood on a past day, and every row carries that day, so a quote in an agreement stays checkable later.

Input

Nothing is required. A run started from the Console uses the prefilled input (query medical device, cfrTitle 21) and writes 50 matching sections of title 21. Prefill is a Console feature: an API call with {} searches the whole code without a phrase and writes the first 50 sections in force, so pass query (and cfrTitle) in API calls.

Reference

Modes

structure and versions read one title, so they are the two modes that need cfrTitle. titles and agencies need no input at all — run them once, keep the values, and use them in the filters of the other modes.

Sort orders

structure always follows the code's own order, titles its numbering and agencies its alphabet.

CFR titles

The 50 titles are the top level of the code. The ones people ask for most:

Run { "mode": "titles" } for the full list with names and the day each title is current through — that is also the honest answer to "how fresh is the text I am quoting".

Agency slugs

agencySlugs filters search by the chapters an agency owns, which is broader than one title and narrower than the whole code. A few common ones:

Run { "mode": "agencies" } for all of them with the chapters each one writes and the sub-agencies below it. A full name (Environmental Protection Agency) or another letter case is turned into the slug. A slug the dictionary does not hold stops the run with the closest slugs as suggestions (did you mean "agriculture-department"?) — a near miss is never swapped for a different agency, and never answered with a silent empty run.

How citations and links are built

citation is assembled from the hierarchy of the node: 21 CFR 820.10 for a section, 40 CFR part 141 for a part, 21 CFR part 820 subpart A for a subpart, 40 CFR part 58, Appendix E to Part 58 for an appendix,

21 CFR chapter I, subchapter H

onlyNew watch lists

With onlyNew: true the actor keeps the key of every row it wrote in the key-value store ecfr-regulations-state of your Apify account and writes only rows whose key is not there yet.

Examples

Which codified FDA rules mention a topic

{ "mode": "search", "query": "quality management system", "cfrTitle": 21, "order": "relevance", "maxItems": 20 }

One agency's sections across the whole code, in the order the code prints them

{ "mode": "search", "query": "pfas", "agencySlugs": ["environmental-protection-agency"], "order": "hierarchy", "maxItems": 20 }

A weekly watch list of sections that changed (only what is new since the last run)

{ "mode": "search", "query": "safety", "lookbackDays": 7, "order": "newest", "onlyNew": true, "maxItems": 50 }

The table of contents of one part, for a RAG index

{ "mode": "structure", "cfrTitle": 21, "cfrPart": "820", "nodeTypes": ["subpart", "section"], "maxItems": 40 }

Every recorded amendment of a part, newest first

{ "mode": "versions", "cfrTitle": 40, "cfrPart": "141", "order": "newest", "substantiveOnly": true, "maxItems": 30 }

Where a phrase is regulated, before exporting any sections

{ "mode": "counts", "query": "artificial intelligence", "maxItems": 30 }

How a single section used to read

{ "mode": "search", "cfrTitle": 21, "cfrSection": "820.30", "includeHistory": true, "order": "oldest", "maxItems": 20 }

Output

One row per section version, tree node, amendment or dictionary entry. The first row of a real run with the default input (mode: "search", query: "medical device", cfrTitle: 21, 50 rows in six seconds):

{
  "mode": "search",
  "nodeType": "section",
  "citation": "21 CFR 892.2020",
  "titleNumber": 21,
  "titleName": "Food and Drugs",
  "chapter": "I",
  "subchapter": "H",
  "part": "892",
  "subpart": "B",
  "subjectGroup": null,
  "section": "892.2020",
  "appendix": null,
  "identifier": "892.2020",
  "heading": "Medical image communications device.",
  "hierarchyHeading": "§ 892.2020",
  "chapterName": "Food and Drug Administration, Department of Health and Human Services",
  "subchapterName": "Medical Devices",
  "partName": "Radiology Devices",
  "subpartName": "Diagnostic Devices",
  "sectionName": "Medical image communications device.",
  "excerpt": "Identification. A medical image communications device provides electronic transfer of medical image… image data between medical devices. It may include a physical communications medium, modems… image review software functionality for medical image processing and manipulation, such",
  "score": 51.984,
  "startsOn": "2021-04-19",
  "endsOn": null,
  "inForce": true,
  "changeTypes": ["effective"],
  "structureIndex": 82397,
  "reserved": false,
  "removed": false,
  "asOfDate": "2026-09-24",
  "found": true,
  "url": "https://www.ecfr.gov/current/title-21/chapter-I/subchapter-H/part-892/subpart-B/section-892.2020",
  "fetchedAt": "2026-09-27T20:58:29.940Z"
}

Filled in every mode: mode, nodeType, citation, titleNumber, url, asOfDate, fetchedAt, found (citation, titleNumber and url are null on an agency row that owns no chapter of the code; citation is null on the single found: false row a run writes when nothing matched). mode, nodeType and found stay in every row even when fields narrows it, so a narrowed dataset still tells which mode wrote a row and which row is the found: false one. The found: false row is a dataset row like any other: it counts towards maxItems and is billed as a row.

Dataset views: Sections (search), Title structure, Amendments, Dictionaries & hit counts. A SUMMARY record in the run's key-value store repeats the filters, the resolved asOfDate, whether it was clamped, the total number of matches, the request count and every warning.

Use it from code / agents

curl -X POST "https://api.apify.com/v2/acts/yadroo~ecfr-regulations/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"mode":"search","query":"quality management system","cfrTitle":21,"maxItems":20}'

import { ApifyClient } from 'apify-client';
const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('yadroo/ecfr-regulations').call({ mode: 'versions', cfrTitle: 40, cfrPart: '141', maxItems: 30 });
const { items } = await client.dataset(run.defaultDatasetId).listItems();

from apify_client import ApifyClient
client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("yadroo/ecfr-regulations").call(run_input={"mode": "counts", "query": "artificial intelligence", "maxItems": 30})
items = client.dataset(run["defaultDatasetId"]).list_items().items

MCP: add https://mcp.apify.com to Claude / Cursor / any MCP client and call the yadroo/ecfr-regulations tool with the same JSON input.

For agents: use counts first, search second. A phrase like "artificial intelligence" matches sections in a dozen titles. One counts run costs a few dozen rows and tells the agent which titles and parts to look in; the follow-up search then carries cfrTitle and cfrPart and pays for the sections that matter instead of paging blindly. Pass fields to trim the row to what the model actually reads — ["citation","sectionName","excerpt","url"] is usually enough.

Pricing

Pay per event: $0.001 per run start + $0.002 per dataset row. The start event is charged on every run. A run that matches nothing writes one found: false row saying so, and that row is billed as a row ($0.002); an onlyNew run whose matches were all delivered before writes no row and costs only the start. Bronze −10 %, Silver −20 %, Gold and above −30 % on the row price; the start event is the same on every plan; platform usage is included.

Your maximum cost per run is respected: the actor reads how many rows it still pays for, writes exactly that many and ends with "Stopped at your spending limit: N rows delivered" — never a row it cannot charge, never a silent cut.

There is no separate compute charge. The runs behind this README finished in a few seconds each — no browser; a search is two requests, a part's structure three. The actor asks for 512 MB because its heaviest job parses a whole title's tree: title 40 is 9.4 MB of JSON and 30 311 nodes, and walking every one of them needs well under half that memory. Passing cfrPart or cfrChapter cuts the subtree before it is walked, which is why exporting one part is the cheap case.

Limits & FAQ

• No full section text, and that is the source's rule, not a shortcut. robots.txt of the eCFR disallows the two endpoints that serve rendered and raw section text for every crawler, so this actor does not read them. You get the citation, the full hierarchy, the headings, the matched excerpt, the version window, the amendment dates and the link to the official page — open the link (or quote the citation) when you need the wording itself.
• The code is one to three days behind today. The source publishes the code up to a processed day, which every row reports as asOfDate; mode: "titles" shows it per title as upToDateAsOf. A date the source has not reached is clamped down to the newest available day, logged as a warning and flagged in SUMMARY.asOfDateClamped — it is not an error.
• 10 000 hits per query. The search index refuses to page past that offset. A run that reaches the ceiling says so in the status message and in SUMMARY.resultWindowHit; narrow it with cfrTitle, cfrPart, agencySlugs or a change window.
• Only the JSON API is read. The site's HTML pages answer a redirect to a different domain when they are requested from a data-centre address, so this actor never fetches a page — it talks to www.ecfr.gov/api only. Links in the rows are for you to open in a browser.
• A missing node gives a row, not silence. A part or section that does not exist on the chosen day, a reserved title, or a filter combination with no data yields exactly one row with found: false and a heading that says what was missing. A section that a rewrite removed (21 CFR 820.30 after the February 2026 rewrite of part 820) has no current page, but mode: "versions" still has its amendment history.
• Superseded versions need includeHistory. By default a section appears once, as it stood on asOfDate. With includeHistory: true you get one row per version, each with its own startsOn, endsOn and inForce — which is how you diff a section over time.
• inForce is computed, not copied. A dated hit can still carry a non-null endsOn when a later version already exists, so "in force" is worked out from startsOn, endsOn and asOfDate rather than read off the source.
• No published rate limit, so the actor stays polite. Requests are sequential with a short pause, 429 and 5xx answers (the amendment log occasionally returns 502) are retried with backoff, and a run makes at most 40 requests. A search writes each page as it arrives, so if the cap (or a source failure after the first rows) ever stops it, the rows already read are kept and the status says why and what to narrow. An onlyNew search reads 1000 hits per page, so even a watch that has to page past thousands of delivered rows needs at most ten pages.
• Spending limit and timeout. The run stops at the last row your maximum cost per run pays for ("Stopped at your spending limit: N rows delivered") and stops starting requests shortly before its timeout ("Stopped before the run timeout: N rows saved") — both end as a succeeded run with the rows written so far.
• Empty fields are the source's. subpart, subchapter and subjectGroup exist only where the code uses them; excerpt and score only in search; sizeBytes and descendantRange only in structure; one agency in the dictionary owns no chapter of the code at all.
• onlyNew keeps one list per filter combination. Changing the query, the title or the part starts a new list, so the first run of a changed watch writes everything it finds — details in onlyNew watch lists.
• Licensing. The Code of Federal Regulations is a work of the US government and in the public domain; the actor adds the structure, the citations and the links.

Made by Yadroo. Sibling actors: federal-register-documents (the daily rules and notices behind these amendments), sec-edgar-filings, openfda-records, cisa-kev-vulnerabilities, bls-release-calendar.