This register records exactly one declared alternative to ghost-cursor: HumanCursor, which reads NRND at the last field check. ghost-cursor itself reads ACTIVE, and 1 of the candidates carry a state mark other than ACTIVE or SEMI, which is the part a list without check dates cannot tell you. Every row is judged by the same rubric, from the same sweep, on the same fields.
Every row below is judged by the same rule as ghost-cursor itself, from the
same field check. Supersession is only printed where a project or its
author says so — this register does not infer that one tool replaces
another because it is newer.
Read the state column with the licence column. A permissively licensed
project that stopped shipping is not safer than a copyleft one that did
not.
What ghost-cursor does, so you know what you are replacing
Generates Bézier-curve mouse paths with human-plausible acceleration for Puppeteer, addressing the behavioural layer rather than the static fingerprint one.
A substitute has to do that job, not merely occupy the same category. Behavioural signals are only worth investing in once the static ones are consistent; a perfect mouse path on a contradictory fingerprint does not help. Check whether your target actually collects pointer telemetry before adding the dependency.
How this register decides what counts as an alternative
HumanCursor is recorded as a substitute for the same job. The alternative relation is symmetric and carries no ordering at all: it says these things answer the same question, and it says nothing whatsoever about which answers it better.
Nothing on this page is a ranking, a score or a recommendation, and nothing on it is a detection-test result — no entry in this register has been run against any anti-bot system by this register. What the table gives you is four measured columns and one derived one, applied identically to every row.
How to read the comparison above
Not one of the candidates is currently shipping. That is the single most useful thing this page can tell you, and it is the thing no undated list can: the whole neighbourhood around ghost-cursor has gone quiet, and replacing it means either adopting something equally stale or leaving the approach entirely.
Every entry in the table declares a licence, which is less common in this catalogue than it should be — the entries that declare none are grouped separately. Read the licence column beside the state column rather than after it: a permissively licensed project that stopped shipping is not a safer dependency than a copyleft one that did not.
Why NRND: pushed_at = 2024-11-28 — 624 days, inside the 365–1095 day NRND band.
Field
Value at this printing
pushed_at
2024-11-28
archived
false
latest_release
—
license.spdx_id
MIT
stargazers_count
503
http_status
200
fetched_at
2026-08-14T09:12:00Z
rubric_edition
2nd, in force
Endpoint: GET https://api.github.com/repos/riflosnake/HumanCursor
unauthenticated · 60 req/hr
A live read does not change what is printed above it. The site records that you ran it; what GitHub returned is not kept. The printed state moves only when the register is swept and a new printing is issued. raw JSON
Answered from the same fields the table above prints, and from no other source.
What is the best alternative to ghost-cursor?
This register does not answer that, because it publishes no ranking and has tested no tool to rank. What it can tell you is which candidates are still shipping: none of them, at this field check. The rest of the decision is fit, licence and how much of your own stack you are prepared to change.
Are these ghost-cursor alternatives still maintained?
HumanCursor reads NRND. Each of those states is derived from that entry's own measured fields by the rubric in force, at the field check printed in the folio, and the clause that produced it is printed on the entry page.
Why does this list not rank the alternatives?
Because a ranking would be an opinion wearing a number's clothes. Every figure on this site is a date, a count, an SPDX identifier or an HTTP status you can re-fetch from the endpoint named beside it, and there is no honest way to compress those into an ordering without inventing weights nobody published. The rubric sets states; it does not rate them against each other.
Where does the alternatives data for ghost-cursor come from?
From declared edges in the register itself, never inferred. Each edge is stored once and reversed at build time, and every field in the table is read from the same sweep that produced the rest of the catalogue. The whole set is downloadable as CC0 JSON and CSV, including the supersedes, superseded_by and alternative_to columns.
Everything in section 02 is a patch applied to somebody else's automation stack, which means every one of them inherits that stack's release cadence and breaks on its own schedule. Donut Browser starts from the other end: the fingerprint handling is in the browser it ships rather than in a layer bolted onto a driver, and the local REST API and MCP server mean automation talks to it directly instead of through a patched client library.