USDA Local Food Directory: Farmers Markets, CSAs, Food Hubs

USDA's bulk export is free; you pay for the work already done on it. One schema across five directories (farmers markets, CSAs, food hubs, on-farm markets, agritourism: about 28,000 listings), addresses parsed into street, city, state and ZIP with the ZIPs USDA left out filled from the US Census Geocoder (98.6% of Texas farmers markets and CSAs have one), numeric coordinates, real lists, duplicates removed, state, keyword and last-update filters, a schedule to keep it fresh, and you pay only for the rows you keep: $2.00 per 1,000 listings, about $56 for the whole country.

Every row also has the last-update date, the products, production methods (organic, no pesticides, ...), accepted payments and food-assistance programs (SNAP, WIC) the listing reports, and a link to the listing's page on the USDA portal. Use it for a local-food map, a CRM list or a research dataset.

No login, no proxy, no browser. The Actor reads the portal's public bulk-export files (one request per directory) and, for addresses USDA publishes without a ZIP, asks the free US Census Geocoder for it (see fillMissingZip), so a full run of all five directories takes about two to three minutes and a default run about half a minute.

What you get

• Five directories in one schema: farmersmarket, csa, foodhub, onfarmmarket, agritourism.
• Parsed addresses: USDA publishes one free-text address per listing in many shapes. The Actor reads street, city, state (two-letter code) and zip from it and always keeps the original text in address. Anything it cannot read with confidence is null, never guessed.
• ZIP codes USDA left out, filled in: about 6% of addresses (11% in Texas) have a street, city and state but no ZIP (10401 Scott Street, Houston, TX, USA). The Actor looks those up in the official, free US Census Geocoder by street, city and state (never by coordinates) and sets zip with zipSource: "census-geocoder". A ZIP USDA published stays zipSource: "usda"; no ZIP is null.
• Clean values: coordinates as numbers (blank and placeholder values become null), update dates as YYYY-MM-DD, the portal's ;-joined answers as real lists, and long text with line breaks tidied.
• Filters: states, keywords (name, description, city, products, production methods) and "updated on or after".
• Polite by design: robots.txt is read on every run, requests are spaced, failures are retried with backoff.

Use from AI agents (Apify MCP)

Add the Actor to an MCP client with https://mcp.apify.com/?actors=flipvo/usda-local-food-directory. It appears as one tool, flipvo--usda-local-food-directory, whose arguments are the fields below. Nothing is required: a call with no arguments reads the farmers-market directory. Pass directories and states (lists) to narrow it. The result names the run's dataset; fetch the rows with the dataset tool and ask for the fields you need, for example name,city,state,zip,latitude,longitude,updatedOn,listingUrl.

Input

Example: organic-minded Texas and Oklahoma CSAs and farmers markets updated since 2024:

{
    "directories": ["farmersmarket", "csa"],
    "states": ["TX", "OK"],
    "keywords": ["organic"],
    "updatedAfter": "2024-01-01",
    "maxItems": 5000
}

Output

One dataset row per listing. docs/sample-output.json holds one real row from each directory from a real run, plus one whose ZIP came from the Census Geocoder; a row looks like this:

{
    "directory": "csa",
    "listingId": "202117",
    "name": "Tasty Little Leaves Microgreens, LLC",
    "address": "Jacksonville, FL 32246, USA",
    "street": null,
    "city": "Jacksonville",
    "state": "FL",
    "zip": "32246",
    "zipSource": "usda",
    "latitude": 30.309322,
    "longitude": -81.504777,
    "updatedOn": "2026-07-23",
    "products": ["Microgreens"],
    "seasonMonths": "Year-round",
    "productionMethods": ["Non-Certified, but Practicing Organic", "Non-GMO", "No pesticides"],
    "listingUrl": "https://www.usdalocalfoodportal.com/fe/flisting/?directory_type=csa&lid=202117",
    "scrapedAt": "2026-11-04T20:05:25.869Z"
}

Each directory has the columns USDA gives it, and a field a directory has no column for is left out of its rows rather than delivered empty: products and seasonMonths for CSAs and food hubs, shareSizes for CSAs, facilities for on-farm markets and agritourism, services for food hubs and agritourism, venue for farmers markets, legalStatus and yearEstablished for food hubs, deliveryMethods for every directory but food hubs. The farmers market export has no product list, season or opening hours, so farmers market rows carry none. A field the directory has but the listing left blank is null or []. zipSource says where zip comes from: usda (written in the address), census-geocoder (matched in the US Census Geocoder, see below) or null (no ZIP). A listing's zip is null only where USDA lists none and the Census Geocoder cannot match the street in that city (for example a highway without a house number); it is never guessed and never derived from coordinates. Export as JSON, CSV, Excel or via the API. A per-directory summary (rows in the export, skippedUnreadable, skippedDuplicate, matched, delivered, error) is saved to the SUMMARY record of the run's key-value store.

A real run of all five directories on 2026-11-04 delivered 27,956 listings: 7,184 farmers markets, 2,008 CSAs, 483 food hubs, 4,707 on-farm markets and 13,574 agritourism farms. Measured on that data, the state was readable for 99% of listings, the city for 97%, the ZIP for 96% (92% from USDA's addresses alone, the rest from the Census Geocoder), coordinates for 99% and a last-update date for 86%. In Texas farmers markets and CSAs (285 listings) the city was read for every listing and the ZIP for 98.6% (253 from USDA's addresses, 28 from the Census Geocoder; 88.8% before the geocoder).

Address accuracy, measured

Filled is not the same as right, so the parsed fields were checked too. On 2026-11-05 (05:03Z to 05:04Z UTC), 40 listings were drawn at random (seed 42) from each of the five directories of that day's export (27,956 listings). Each street, city, state and zip was checked by hand against USDA's own address text. Every ZIP filled from the Census Geocoder was checked against a second, independent Census answer: the ZIP Code Tabulation Area at USDA's own coordinates.

• Zero wrong values in 200 puts the true wrong-value rate below about 1.5% (95% confidence, the rule of three).
• The fix this check found: a unit or a landmark written before the city with no comma ("2nd Floor Great Barrington", "behind Ritz Theater Newburgh") became part of the city. It is now cut off into street, but only when the words before the city are plainly not a place name (a number, "Suite", "lot", "behind" and the like) and the rest is a Census place. Across the whole export this changed 54 listings, and every one of them got the right city.
• Left null: a lone "Chester" with no state, and "01072, Shutesbury, Massachu" (the ZIP first and the state cut short). The full address is still delivered.
• Checked against the text as written: "Idaho Falls, IA" is delivered as IA (Iowa), because that is what USDA wrote.

The sample, with each listing's link, USDA's text, our answer and the hand-checked answer, is kept with the Actor's source (docs/address-accuracy-2026-11.json), next to the scripts that draw, re-parse and score it. A test fails the build if this table stops matching that file. Ask in the Issues tab for a copy.

Pricing

Pay per event: $2.00 per 1,000 listings delivered ($0.002 per listing). Listings you filter out, duplicate rows and unreadable rows are never charged. Apify also charges its standard start event, $0.00005 per run. Platform usage is not charged on top. maxItems caps the bill: $2.00 at the default of 1,000. All five directories for Texas came to 892 listings on 2026-11-04, about $1.78; the whole country (27,956 listings) is about $55.91 for a full refresh.

The five directories in one schema, the address work, the cleaned values, the real lists, the removed duplicates and the filters are described above; if you only need one raw file once, download it from USDA at usdalocalfoodportal.com.

Good to know

• No phone numbers, emails or contact names. The public bulk export does not contain them, so this Actor does not return them. It does not open listing pages or look anywhere else for them.
• The data is what listings reported to USDA, and many listings are old: the updatedOn date tells you how old. Use updatedAfter to keep only recently confirmed ones. Listings with no date in the export have updatedOn: null.
• The export has no separate state, city or ZIP columns. They are read from the address text (state names and codes, ZIP and ZIP+4, a missing comma before the city, ZIPs that lost their leading zero), so about 1% of listings have no state and about 3% no city (the full address is still there). address is always USDA's text; a ZIP from the Census Geocoder is added to zip only, not to address. When you filter by states, listings with no readable state are left out.
• A listing id that appears twice in one directory's export is delivered once and counted in skippedDuplicate; rows without an id or name are counted in skippedUnreadable. Neither is charged.
• The same farm can appear in more than one directory (for example a CSA that also runs a farm stand). Listing ids are unique within a directory, so use directory + listingId as the key.
• The portal sometimes answers with an error page instead of data. The Actor treats that as a failed directory, reports it in SUMMARY and still delivers the other directories; if every directory fails the run fails.
• A run keeps the whole export of one directory in memory (the largest is about 40 MB); 512 MB of run memory is enough.

ZIP codes from the Census Geocoder

With fillMissingZip on (the default), every delivered row that has a street, a city and a state but no ZIP is sent to the US Census Geocoder (locations/addressbatch, benchmark Public_AR_Current, up to 250 addresses per request; the geocoder's documented limit is 10,000). It is a free public service of the US Census Bureau with no robots.txt (so nothing is disallowed), and the Actor sends the same self-identifying User-Agent as for the portal. A ZIP is taken from a match only when:

• the match type is Exact or Non_Exact (a tie or no match gives nothing), and
• the matched state and city are the ones in USDA's address (spelling variants such as Lee's Summit and LEES SUMMIT, Mt and Mount, Jackson Township and Jackson count as the same).

The city check matters for Non_Exact matches, where the geocoder adjusts the street: "Between 4th and 5th Avenue, Garnett, KS" came back as an intersection in Lenora, KS, and is refused. Only the ZIP in the matched address text is used; the coordinates the geocoder returns are ignored.

How right is it? On 2026-11-04 the ZIP of 3,000 addresses that USDA does publish with a ZIP was withheld and looked up the same way: of the 2,136 that passed the rule above, 96% came back with the same ZIP USDA published (97% of the Exact ones, 93% of the Non_Exact ones), and 99% with the same first three digits; the others are usually the neighbouring ZIP of the same town, because the geocoder reads the ZIP from Census address ranges, not from the postal service's delivery file. A spot check of 20 geocoded Texas addresses against a second Census call (the one-line address endpoint) and against the ZIP area at USDA's own coordinates found the same ZIP for 20 of 20 and 19 of 19 readable (the area lookup returned nothing for one). The geocoder is best effort: it is sometimes slow, so a run may wait up to about 75 seconds for one batch; if a batch cannot be answered the run logs a warning, those rows keep zip: null, after two failed batches in a row it stops asking, and the run never fails because of it. Turn fillMissingZip off to get USDA's addresses exactly as published.

Legal and data note

This Actor reads the USDA Local Food Portal's public bulk export, federal data that USDA Agricultural Marketing Service publishes for anyone to download, through the portal's public download endpoints. It does not log in, bypass any protection or use proxies, and it checks each host's robots.txt before every run (neither the portal nor the Census Geocoder publishes one, so nothing is disallowed). The portal rejects requests whose user agent does not start with Mozilla/5.0; the Actor sends the standard self-identifying bot form,

Mozilla/5.0 (compatible; usda-local-food-directory/1.0; Apify Actor; public USDA data; +https://apify.com)

The Actor keeps only what USDA publishes in that export. Listings are businesses and organizations (market managers, farms, hubs). The export carries no phone, email or contact-name fields, and the Actor returns none; free text that a listing owner chose to put in the public description or location note is passed through unchanged, as USDA publishes it, and is not extracted, enriched or combined with any other source. You are responsible for using the data in line with the law, including data-protection and marketing rules (for example CAN-SPAM, TCPA, GDPR where EU residents are concerned). This Actor is unofficial and not affiliated with, endorsed by or sponsored by the USDA.

Support

Found a listing that parses wrongly or a field that is missing? Open an issue on the Actor page and include the directory and listingId.