Splash is unmaintained, and it was never a real browser

Splash is a QtWebKit render service whose default branch was last pushed in August 2024, and QtWebKit is an engine no consumer browser ships. Both facts matter, but the second is the one no spoof reaches: every JavaScript-visible surface reports an engine that matches no real audience, and that mismatch lives below any header, so no user-agent string or page-level plugin corrects it.

Splash is a render service: you hand it a URL over an HTTP API, it runs the page and hands back HTML, a screenshot or a rendered DOM. For a long time that was a reasonable way to execute JavaScript on a server without driving a full browser yourself. It is still packaged, still starred, and still recommended in old answers.

This page is about two separate problems with reaching for it in 2026. The first is that it has stopped moving. The second is deeper and would matter even if it had not: the engine it renders with is one that no person browses the web with, so the identity it presents matches no real population before you have set a single header.

What Splash actually is

Splash renders pages with QtWebKit, the WebKit binding that ships with the Qt5 toolkit. That is not a detail of how it is built, it is the whole product: a scriptable WebKit executed headless and exposed over HTTP.

The consequence is that everything a page can read in JavaScript comes from QtWebKit. The navigator.userAgent string, navigator.vendor, the way the canvas rasterizes a glyph, the WebGL renderer string and the WebGL pixels behind it, the set of supported codecs, the exact rounding of a layout measurement: all of it is QtWebKit’s, because there is no other engine underneath to borrow from.

QtWebKit is an engine nobody browses with

QtWebKit is a WebKit fork that no consumer browser ships today, so the real-world audience running it is zero, not merely small. This is the part that a user-agent string hides and does not fix.

Every browser a real visitor uses is one of three engines: Blink (Chrome and the Chromium family), WebKit proper (Safari), or Gecko (Firefox). QtWebKit is a fourth lineage. It forked from WebKit years ago, and it is not the WebKit that ships in any consumer browser today. When you measure a real audience, the count of visitors on QtWebKit is zero, not small.

So a detector does not need a cleverly hidden tell. The engine is the tell. A canvas hash that no Blink, WebKit or Gecko build has ever produced, a WebGL renderer string that pairs with no real driver, a set of feature-detection results that line up with no shipped browser: these are not spoofable properties sitting on top of a normal engine. They are what the engine genuinely is. You can set navigator.userAgent to a current Chrome string, and the first canvas or WebGL read disagrees with it, because the pixels come from somewhere the header cannot touch. This is the same failure mode as a renderer string that says NVIDIA while the pixels say software, except here the entire engine is the mismatch rather than one field of it.

It is also why “Splash plus a stealth plugin” does not close the gap. A page-level plugin rewrites the answers JavaScript gives to a handful of known questions. It does not change which engine drew the canvas or handed back the WebGL renderer, and those are exactly the surfaces that give a non-consumer engine away.

The maintenance date, which is the easy half

Checked at the source rather than from memory: the repository’s last push to its default branch is dated 2024-08-02. As of this writing that is very close to two years with no commit, on a project of roughly 4,200 stars. The repository is not archived, so nothing announces the state on the page. You have to read the date.

Two years is a long time for a browser engine specifically, because the population it has to blend into moves every few weeks. New engine versions ship, feature detection shifts, the shape of a normal fingerprint drifts. An engine frozen in 2024 does not drift with it. But the staleness is the smaller problem here. A frozen Blink would at least be a frozen version of an engine real people ran. A frozen QtWebKit is a frozen version of an engine they never ran at all.

Why a header or user-agent spoof cannot reach it

A header or user-agent spoof cannot reach QtWebKit’s tells because it rewrites the request, not the engine, and the canvas, WebGL renderer and feature-detection results are computed by the engine. It helps to separate the two layers, because people spend effort on the wrong one.

The layer you can rewrite from outside the engine is the request and a few named JavaScript properties: the user-agent header, the Accept-Language header, a pinned navigator.platform. These are cheap to change and cheap to check against each other.

The layer you cannot rewrite from outside is what the engine computes: how it draws, how it measures, what its graphics stack reports, how it answers hundreds of small capability questions. Those come from the compiled engine. To change them you change the engine, and if you are running a render service the engine is the service.

Layer What it includes Can a spoof change it?
Request and named JS properties user-agent header, Accept-Language, navigator.platform Yes: cheap to set, cheap to cross-check
What the engine computes canvas pixels, WebGL renderer, layout rounding, feature detection No: only changing the engine changes these

This is why the durable difference between a render service and a real browser is not “how new is it”. It is which of these two layers you are standing on. A tool built on a consumer engine starts from a body that a real audience actually has and works on the handful of outside-the-engine tells. A QtWebKit service asks you to defend the one layer you cannot reach.

What a real engine changes, with code

The alternative is to drive a real consumer engine with stock automation, so the answers below the header come from an engine a real audience runs, and the seed makes them reproducible. invisible_playwright launches a Firefox patched at the C++ level and returns an ordinary Playwright Browser, so every method is the one you already know.

Read the surfaces that a render service cannot honestly produce, straight off the page. The renderer name below comes from the WEBGL_debug_renderer_info extension, which every engine answers from its own graphics stack rather than from a header:

from invisible_playwright import InvisiblePlaywright

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com")

    ua = page.evaluate("() => navigator.userAgent")
    vendor = page.evaluate("() => navigator.vendor")

    renderer = page.evaluate(
        """() => {
            const gl = document.createElement('canvas').getContext('webgl');
            const ext = gl.getExtension('WEBGL_debug_renderer_info');
            return ext ? gl.getParameter(ext.UNMASKED_RENDERER_WEBGL) : null;
        }"""
    )

    print("userAgent:", ua)
    print("vendor:   ", vendor)
    print("renderer: ", renderer)

Because seed=42 fixes the identity, the userAgent, the WebGL renderer and the canvas hash come back identical on every run, so a failing page is reproducible instead of a new draw each time. That reproducibility is the point of a seed and is covered in the quickstart.

To confirm the engine claim is honest rather than a header, read the same value twice in one session and compare it to a stock browser on the same machine. A render service gives itself away on the second read or on the diff, never on the header alone:

from invisible_playwright import InvisiblePlaywright

def canvas_hash(page):
    return page.evaluate(
        """() => {
            const c = document.createElement('canvas');
            const ctx = c.getContext('2d');
            ctx.textBaseline = 'top';
            ctx.font = '14px Arial';
            ctx.fillText('the quick brown fox', 2, 2);
            return c.toDataURL();
        }"""
    )

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com")
    first = canvas_hash(page)
    second = canvas_hash(page)
    assert first == second, "a stable canvas is the cheapest tampering check there is"
    print("canvas stable:", first[:48], "...")

The method for reading these reports instead of trusting a verdict is its own page: how to test whether your browser is detected.

One honest caveat

Running a real engine fixes the engine layer, not the network layer. The exit IP is not a browser property and no engine touches it: a genuine Firefox on a datacenter address that a page has already seen a thousand times today still loses, and that is a different fight from the one this page is about. Work it in order rather than buying a proxy on day one, which is the whole point of the detected-on-one-site checklist.

Conclusion

Splash’s problem is not only that its last push is dated 2024-08-02. It is that the engine underneath it, QtWebKit, is one no consumer browser ships, so every JavaScript-visible surface reports a render-service engine that matches no real population. That mismatch lives below the request, so no user-agent string, Accept-Language header or page-level plugin reaches it. If you need a page executed on a server and you need it to look like a person’s browser, the engine has to be one a person actually runs.

Short answers to the questions that lead here

Is Splash still maintained? Its default branch was last pushed on 2024-08-02, close to two years ago at this writing, and the repository is not archived so nothing on the page announces it. Read the date yourself before relying on it.

Can I make Splash pass fingerprint checks with a user-agent override? No. The user-agent is a header. The canvas, the WebGL renderer and the engine’s feature detection come from QtWebKit, and they contradict any header you set.

What is wrong with QtWebKit specifically? It is a WebKit fork that no consumer browser ships today, so the audience running it is zero, not small. Real visitors are on Blink, WebKit proper or Gecko.

Does adding a stealth plugin to Splash fix it? A plugin rewrites a few named JavaScript properties. It does not change which engine drew the canvas or reported the GPU, which are the surfaces that give a non-consumer engine away.

Is a render service the same as a headless browser? Not for this purpose. A render service exposes an engine over HTTP; if that engine is not one real people browse with, the output looks like the service no matter how you call it.

If I switch to a real engine, am I done? For the engine layer, yes. The exit IP is separate and no engine changes it, so a real browser on a burned datacenter address can still be blocked.

Sources

  • The Splash repository, read at its own source: default-branch last push dated 2024-08-02, roughly 4,200 stars, not archived, described as a scriptable browser as a service rendering with the Qt5 toolkit.
  • This project’s own testing notes on why an engine-level surface cannot be reached by a header or a page-level override, and why a stable value read twice is the cheapest tampering check.

See also: Chromium is not Chrome, and detectors know the difference for the same engine-lineage problem on the automation side, and how to scrape a site that blocks headless browsers for the machine-level tells that survive every header change.

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. The engine underneath is the one thing a header cannot rewrite, which is the whole argument of this page.

Back to top

MIT licensed. Every page here is written against the current source of the thing it describes, and several record something we got wrong first.

This site uses Just the Docs, a documentation theme for Jekyll.