Wellfound is what AngelList Talent became: the job side of AngelList, where startups post vacancies with a pitch, a size band and — often — a salary range and an equity range. Getting that out is mostly a matter of understanding how the site addresses its own listings.
Wellfound filters by path, and that constrains the input
Filters live in the URL path rather than in a query string. A role is a path segment, a location is a path segment, and the combinations that exist are the ones the site has routes for. That is not a detail: it is why one filter combination cannot be requested at all.
Step 1 — choose roles or categories
Wellfound publishes 518 role slugs across 13 categories. Name them individually in roles, or name a whole category in roleCategories and let the Actor expand it. Engineering alone holds 227 roles, so a category run is a long one.
A role yields about 1,950 unique jobs on average, and one listing request returns about 28 of them.
Step 2 — add locations, or remote, but not both
There are 306 location slugs, and each role is crossed with each location — so three roles and four locations is twelve sweeps, not seven. Leave locations empty for no location filter, which is also the only way remoteOnly takes effect.
Step 3 — cap the run with the right knob
maxJobs is the cap that matters, because delivered rows are what you pay for. maxPagesPerRole looks like the equivalent and is not: measured ceilings run from 37 pages for office-manager to 130 for engineer, so one value gives you a whole corpus in one role and a third of it in another — the measurement.
maxConcurrency defaults to 6 and caps at 16, which is the highest level measured without a block — where the rate-limit ceiling is.
Step 4 — run it
{
"roles": ["software-engineer", "product-manager"],
"locations": ["san-francisco", "new-york"],
"maxJobs": 500,
"enrichFromJobPage": true,
"emitCompanies": true,
"maxCacheAgeDays": 1
}enrichFromJobPage adds structured salary, coordinates, benefits and the company's geocoded headquarters at one extra request per job. The full description arrives without it.
Reading the output
58 fields per job. The description is at 100% of 2,229 measured jobs, location names at 90.5%, the raw compensation string at 79.5%, parsed salary at 86.7% of enriched rows and coordinates at 93.3% of enriched rows. The full field reference has all of them.
Two fields are published as ranges rather than averages, because their category spread is wider than a mean can describe: ats_source at 21–83% and years_experience_min at 12.8–75% — why the average is wrong.
What a run costs
$3.00 per 1,000 jobs, $1.50 per 1,000 deduplicated companies, $0.01 per 1,000 starts. Five hundred jobs with companies is about $1.60. The description and the detail-page enrichment are in the job price and are never billed separately, so a run's cost is the row count times the row price with nothing hidden.
Four mistakes that waste a run
Setting remote and a location together. The site has no route for it, so the remote flag is dropped. Run them as separate jobs if you need both.
Capping on pages instead of rows. The same page cap means different coverage in every role, and nothing in the output says which you got.
Running from a datacentre IP. 24 of 36 requests came back as HTTP 429 where residential returned 48 of 48 as 200.
Treating `scraped_at` as freshness. Cloudflare's cache sat in front of job detail pages at up to 23.4 hours, and data_age_hours counts it — why the two differ.
