Find someone's public profiles from a username, or a list of them, across all 999 sites the scanner knows, every run. Paste a handle, an @handle or a profile link. Filter by category, country or named sites. Every hit ranked high, medium or low confidence with page details. No login, no API key.
[1.2.0] - 2026-10-07: runs that used to fail now answer
Apify flagged the Actor for reliability: 23 of 211 runs failed in the 30 days to 2026-10-07. 13 of them came before
1.1.0; the other 10 were made by users who do not share their runs, so every path in the code that could fail a run
with something to search was closed, and each one was run locally with the real SDK before release.
Changed
A filter that would leave no site to check is left out instead of failing the run. 208 of the 299 category and
country pairs in the form select no site (adult and dating sites based in India, gaming sites based in Angola), and
those runs ended FAILED with "Your filters match no sites". Now the country is left out, the category's sites are
checked, and the message says where they are based. The same holds for named sites none of which is in the list and
for a country no site is based in: options only narrow.
A full name is searched as a handle. "Lisa Vor" was skipped, and on its own failed the run, because usernames
have no spaces. It is now read as lisavor (two to four words of letters, accents dropped), and the summary names
lisa.vor and lisa_vor to add. One name costs what one username costs.
Scan steps fit the run's memory. Each step is its own scanner process of about 300 MB (more when CPU is
short); 8 run at once at the default 4 GB, 5 at 2 GB, 1 at 1 GB. Measured on Apify: the old 8 at 2 GB peaked at
2046 MB and lost two steps to the OOM killer; 5 at 2 GB peaked at 1907 to 1942 MB and finished in 113 s where 4
took 189 s side by side; 2 at 1 GB lost one step, 1 at a time lost none. A step that comes back empty is retried
alone after the others. Sites still not checked are counted in sitesNotChecked, with the reason in the message,
instead of disappearing from the totals.
A spending limit that cannot pay for one profile ends the run as succeeded, with nothing searched or charged.
The message gives both numbers ("this run's spending limit is $0.003, which cannot pay for one profile ($0.005)")
and OUTPUT says not_run. It used to run a full scan it could not deliver. A run timeout too short for one scan
is treated the same way.
An unexpected error no longer loses the run. Rows already saved stay, the summary and OUTPUT are written, and
the message names the error and any usernames that never started. The run fails only when nothing was saved.
Every failed run writes OUTPUT with the reason (a form with no usable username, a build without the scanner).
Rows the spending limit cut are no longer counted as found: the totals match what was delivered and charged.
Filter values of the wrong type are ignored instead of crashing the run; input that is not an object reads as no
input.
[1.1.0] - 2026-09-13
Driven by the Debugging data of the week of 2026-09-05: 11 of 45 user runs failed, and
the causes were not what the success rate suggested.
Fixed
Runs started through Apify's AI channel (meta.origin = APIFY_AI) no longer fail at the
end. SDK 3.x validates the run object against an enum that does not know that origin and
raised inside Actor.set_status_message after all rows were pushed; 8 finished runs a
week were marked FAILED. The status call is now wrapped and can never fail the run.
Several usernames in one run are searched one by one. The scanner treats "alice bob" as a
single handle containing a space and reported instagram.com/alice bob as found; a
six-username comma list from a real user returned nothing for that reason.
Email addresses typed into the username box (3 failed runs from one user in 5 days) are
no longer an error: the part before the @ is searched and the summary points to
holehe-email-osint for the address itself.
Category and country filters work again. The scanner's own --type and --countries
options select zero sites in version 0.45, so the site list is built here from the
scanner's sites.json on every run and passed down as --websites with exact URL
patterns (a bare host token would over-match: t.me also selects about.me).
slow and special scan modes removed from the form: both return no data in the current
scanner. A legacy mode value in API input is accepted and noted, and the fast scan runs.
Time limit is respected: usernames that cannot start before the limit are listed in the
summary instead of the platform killing the run with nothing written.
Added
Progress and chunked delivery. Each username is scanned in steps of about 250 sites,
two at a time; after every step the rows are pushed, the run's status line is updated
with the percentage and counts, and the OUTPUT record carries status: running and
percentDone. A run cut off by the time or spending limit keeps what was delivered.
Charging moved from the platform's default dataset-item event to a custom profile
event at the same $0.005, charged per step as rows land; the summary row is free.
Requires the pricing record to name the profile event at deploy time.
Every run is the deepest scan the scanner can do: all 999 sites by default (was 100),
every output field it can emit (status, type, country, language, rank, metadata,
extracted), 60 parallel workers through src/scan.py instead of the hardcoded 15
(all sites in about 2 to 3 minutes instead of 7), and its permissive detection level.
The scanner's slow/special modes have no code behind them and its "extreme" level
is a stricter gate, so no depth setting is offered: there is nothing deeper to offer.
The confidence filter now works. The scanner's own --filter silently did nothing
unless status was among the requested output fields, so "confident only" returned
50% matches too. Filtering is done here on the match rate: confident = 100%, possible
= 50 to 99%, everything = all. Confidence tiers follow the same thresholds. The default
is "everything": a bare username must miss nothing, and every row is labelled.
A username with nothing else set gets every option at its maximum: all 999 sites,
worldwide, every category, every match, metadata and pattern extraction on, a one-hour
time limit. Every other field only narrows the scan.
Summary counts sites with no profile (sitesNotFound) and sites that did not answer
(sitesFailed), so "0 results" is explained by the numbers.
Rows carry language and metadata (meta tags of confident profiles); the previous
version requested only link, rate, title and text, so the "Add page metadata" option
had no visible effect.
usernames list field for bulk input (up to 25 per run), alongside username.
The site list is built by the actor on every run (top N by popularity from the
scanner's own data), so the number of sites checked is exact and reported. Adult and
dating sites (48) are part of every run, as they always were; the adultSite column
marks those rows and the "Adult and dating sites" category checks only them.
Every profile row carries platform, site, category, categoryDetail, country,
adultSite, siteRank, matchRate and pageTitle; the scanner itself only returns
link, rate, title and text.
One summary row per run with per-username results, the filters applied, notes about
cleaned or skipped input, and the plain-language message also shown as the run status.
The same object is saved as the OUTPUT record for API and agent users.
Named-site input accepts any of the 999 sites by domain or partial name; unknown names
are reported with close matches.
Category dropdown built from the scanner's data, with site counts per category.
Unit tests for input cleaning, site selection, enrichment and the status-message guard
(python3 test_unit.py), plus a two-site real scan when the CLI is installed.
Changed
Input form regrouped into "Who to look up", "Where to look", "Results", "Limits"; every
title and description rewritten in plain words with the technical detail second.
Dataset views show platform, profile URL, confidence, match %, category, site country and
adult flag; the always-empty status, language and rank columns are gone.
trim is always on and no longer a field; method is no longer a field (found profiles
are always what is returned).
siteType values are now category slugs and countries values are country names, both
chosen from dropdowns. A saved task or API call that still sends the old free-text values
(Dating, gb) is rejected by the platform's input validation before the run starts:
pick the category or country again.
[1.0.1] - 2026-06-30
Added
Auto-clean the username input: a pasted @handle, a profile link (twitter.com/elonmusk), or a comma-separated list is now normalized to bare handle(s). Tested live: a raw @elonmusk used to return 21 confidently-broken URLs (github.com/@elonmusk) while still reporting success; it now resolves to the real profiles.
Early input validation: clearly invalid input now fails fast with a plain reason instead of a silent empty or garbage run.
Confidence ranking: every detected profile gets a confidence tier (high / medium / low) from its match rate, and results are sorted strongest-first so the real profiles surface above long-shot guesses.
Clear end-of-run status message: every run reports what it found (with the high-confidence count), and a 0-result or all-weak-matches run says so explicitly instead of looking identical to a strong run.
Changed
Input form regrouped: the username box stands alone on top, with the rest folded into "Scan options" and "Advanced" sections. No options removed.
Rewrote field descriptions and option labels from tool jargon into plain English.
Repositioned the listing for non-developer users (recruiting, people search, fraud, due diligence) while keeping the developer / OSINT / Sherlock-alternative discoverability.
Run summary now reports high / medium / low confidence counts.
Fixed
README input table and output example now match what the actor actually accepts and returns (was listing non-existent usernames, categories, confidence_min fields and a wrong output shape).
Removed em-dashes from the input schema and source.
