Proxies for Legal and Court Records Data Collection

Find your target on the left. That row is your purchase.

Two rows say "buy nothing" and one says "buy a subscription that is not a proxy". That is the shape of this market, because "legal data" hides four workloads: case law is a one-time backfill, dockets are a request-count problem, filed documents are a bandwidth problem at 200 KB to 20 MB each, and adjacent registers are a fragmentation problem.

Check this list before you spend a dollar. Each is first-party and openly licensed, so a proxy in front of it adds latency and nothing more.

Register for a key even where anonymous access works: keyed quotas are higher. A documented rate limit binds to the key, not the address, so adding IPs against a key-scoped quota is money burned. The same trap appears in proxies for government tender data and how to scrape SEC EDGAR filings.

PACER is the most requested target in this category and the one vendors most often oversell.

The federal judiciary charges $0.10 per page for most retrievals, capped at $3.00 per document, with written opinions free and fees waived for accounts at or under $30 in a quarter. Every one of those meters attaches to your login, so the system bills the account regardless of which exit address the request came from. Three consequences follow:

• Rotation gains you nothing. You cannot rotate away from a per-page charge on a credentialed session.
• Aggressive automation risks the account. Bulk retrieval without arrangement with the courts is a terms problem, and terms problems on an authenticated system end in a suspended login.
• Somebody may already have paid. RECAP inside CourtListener holds PACER documents other users bought and contributed, so check it first.

Bills to abolish PACER fees have reached Congress more than once and had not become law as of September 2026, so verify the schedule on pacer.uscourts.gov before you budget. If a vendor pitches residential proxies "for PACER", walk.

The genuine case starts below the federal system, and it is a fragmentation problem rather than an adversarial one. The US has more than 3,000 counties, with clerk and recorder functions run at that level, so a national lien dataset means touching hundreds of independent hosts. Four things make a proxy layer useful:

• Session tokens pinned to source IP. Many legacy Java portals bind the session to the address that created it, so rotating mid-session drops you back on the search form. Know what a sticky session proxy guarantees before you depend on one.
• Per-IP daily search caps. A meaningful minority of county portals cap searches per address per day, and distribution across a pool is the direct fix.
• Country-level geo-fencing. Some portals serve domestic traffic only, and a US exit clears that. State-level targeting is almost never required.
• Politeness across small servers. A county server is not a hyperscale endpoint. Our guide on ethical scraping and rate limiting has the per-host arithmetic.

Notice what is missing: sophisticated bot defence. Most court portals do not run the behavioural fingerprinting you meet on ticketing sites, so the WAF-fronted minority are exceptions to route around, not a reason to size the whole purchase differently.

The decisive variable is the billing model, not the IP class. Run the arithmetic: a lien and litigation pipeline pulling 60,000 filed documents a month at an average 1.4 MB each moves roughly 84 GB. Published residential list prices sat broadly in the low-to-high single digits of dollars per gigabyte as of September 2026 and change often, so price your shortlist from each vendor's current page. Even at the cheap end, 84 GB is a real line item purely for moving bytes.

SparkProxy's published datacenter plans are flat with unlimited bandwidth on every tier: Starter $75/mo, Core $140, Boost $240, Plus $440, all with 30 days validity. The 84 GB costs the same as 8 GB or 800 GB, and for document-heavy collection that one property usually decides the purchase. See understanding datacenter proxy pricing models and residential vs datacenter proxies. The honest counterpoint: if your target list is dominated by portals that hard-block datacenter ASNs, none of that saves you, so test before you commit.

Most buyers oversize. The binding constraint is not your thread count, it is how much each individual court tolerates, so size it as courts swept in parallel multiplied by concurrency per court. Two concurrent requests per host is a defensible ceiling on a county server, so forty courts at two each is eighty threads.

Pro and Pro+ tiers exist in the fair usage policy at 1500 and 2000 threads, quoted rather than publicly priced. Read the speed figures as ceilings: real throughput against a legacy county portal is set by that portal.

The gateway is gateway.sparkproxy.io, with HTTP and HTTPS on port 11000, sticky sessions on 11002 and SOCKS5 on 13000. Point session-bound court workers at 11002 and everything else at 11000. Whitelist slots matter here because compliance teams want a fixed, auditable set of source addresses rather than credentials in a config file. See IP whitelisting for proxies first.

For the WAF-fronted exceptions, the SparkProxy Scraping API bills in credits: 1 for a plain fetch, 5 for a JavaScript render, 10 for a screenshot or PDF. Plans run from Starter at $49 for 250,000 credits a month at 50 concurrent up to Scale at $599 for 8,000,000 at 400, with 1,000 free credits and no card to test a county search flow end to end.

# Fetch a JS-only court search result as markdown, US exit
curl -X GET "https://scrape.sparkproxy.io/api/v1?url=https%3A%2F%2Fexample-county.gov%2Fsearch&render_js=true&format=md&country_code=us" \
  -H "X-API-Key: YOUR_API_KEY"

Which of the two is the better buy is a real fork, worked through in web scraping API vs self-managed proxies.

Send this list. The answers separate serious vendors from resellers.

• Is bandwidth unlimited, and is it in the fair usage policy? For document workloads this is the whole ballgame.
• How long do sticky sessions last, and what happens when one expires mid-flow?
• Is IP whitelist authentication supported, and how many slots?
• What is the logging and retention policy for connection metadata? Get it in writing before procurement asks.
• What is the replacement policy when a subnet gets blocked by a target?
• Is there a real trial? Free credits with no card beat a refund promise, because you run your own targets.

Then test rather than trusting anyone's number, ours included. Push a fixed sample of your twenty most important jurisdictions through each candidate and record status codes, challenge rates and time to first byte per host. Knowing how to detect when your scraper is blocked makes that test meaningful: a 200 carrying a challenge page is the commonest false pass in proxy evaluations.

This shapes your architecture more than any proxy decision.

Downstream use may make you a consumer reporting agency. If court-record output feeds decisions about employment, tenancy, insurance or credit, the Fair Credit Reporting Act applies: accuracy obligations under 15 U.S.C. 1681e(b), extra duties for public records reported for employment purposes under 1681k, provenance per field, and a dispute path. No proxy plan gives you those.

What indexes expose changes under litigation. Federal Rule of Civil Procedure 5.2 requires redaction of Social Security numbers, account numbers, birth dates and minors' names, and most states have an analogue. In California, All of Us or None v. Hanlon (2021) held that Rule of Court 2.507 does not permit date-of-birth search in public case indexes, and courts there withdrew it. Name-only matching is materially less accurate, so design your matcher to survive an identifier disappearing.

Public access is not unrestricted reuse. US law recognises a strong presumption of access to court records, discussed in Nixon v. Warner Communications (1978), but that governs access alone, not portal terms or the privacy law that attaches once you compile a database. Van Buren v. United States (2021) narrowed "exceeds authorized access" under the Computer Fraud and Abuse Act, and the Ninth Circuit in hiQ Labs v. LinkedIn (2022) indicated that scraping public, unauthenticated data is unlikely to violate it. That same dispute later produced a district court finding that hiQ had breached LinkedIn's user agreement: public pages are one risk profile, credentialed systems another, and contracts survive either way.

European judgments are often pseudonymised deliberately, so re-identifying parties by cross-referencing datasets is a GDPR problem you can create with no bad intent. None of this is legal advice. The research-ethics side is covered in using proxies for academic research data collection.

If you need fewer than about twenty jurisdictions and an aggregator such as UniCourt, Trellis or Docket Alarm sells all of them with an SLA, buying beats most honest build estimates: you skip OCR, per-portal maintenance and the constant breakage of legacy search forms. If your differentiator is coverage of the places vendors skip, you are building, and the decision above applies. Either way, ask for a jurisdiction-level coverage list with refresh frequency, not a state map. The map is always green.