Record and replay HTTP traffic with HAR in Playwright

A HAR file is a plain JSON log of every request and response a page made: URLs, headers, status codes and, if you ask for it, the response bodies. Playwright can write one for you while a session runs, and then serve the whole thing back to a later run without touching the network. That second half is the useful part: a test that reads from a frozen HAR is deterministic, offline, and does not depend on a site being up or a proxy being clean.

This page shows the record step, the replay step, both on the patched Firefox this project ships, and then the caveat that decides where the technique belongs. A replayed HAR is a fixture, not a disguise, and the file you captured is not safe to share.

Record every request to a HAR

Point a context at a path with record_har_path and everything it fetches lands in that file when the context closes. The two-line launch is the same one from the Quickstart; the browser it returns is a real Playwright Browser, so new_context and every option it takes work exactly as upstream documents them.

from invisible_playwright import InvisiblePlaywright

with InvisiblePlaywright(seed=42) as browser:
    context = browser.new_context(
        record_har_path="capture.har",
        record_har_mode="full",     # "minimal" drops bodies and timings
        record_har_content="embed",  # response bodies inline in the JSON
    )
    page = context.new_page()
    page.goto("https://example.com")
    page.click("#load-more")
    context.close()  # capture.har is flushed here, not before

record_har_mode="full" keeps timings and sizes; "minimal" keeps only what is needed to replay. record_har_content="embed" writes each body into the JSON, which is what makes the file self-contained and portable. The file does not exist until the context closes, so a crash before context.close() leaves you nothing.

Replay the HAR offline with routeFromHAR

The mirror image is route_from_har. Register it on a fresh context and every matching request is answered from the file instead of the network.

from invisible_playwright import InvisiblePlaywright

with InvisiblePlaywright(seed=42) as browser:
    context = browser.new_context()
    context.route_from_har("capture.har", not_found="abort")
    page = context.new_page()
    page.goto("https://example.com")  # served from capture.har, no network
    print(page.title())
    context.close()

not_found="abort" makes any request the HAR does not contain fail loudly, which is what you want in a test: a silent live fallback would hide the fact that your fixture is incomplete. Use not_found="fallback" only when you deliberately want unmatched requests to hit the real network. There is also an update=True mode on route_from_har that refreshes the file from live traffic instead of serving it, handy for regenerating a capture without rewriting the record step.

The async API is identical in shape; import InvisiblePlaywright from invisible_playwright.async_api and await the context and page calls, exactly as shown in the Quickstart.

Why the same seed matters for a replay test

A HAR freezes the network. It does nothing for the browser identity, and a test is only fully deterministic when both halves are pinned. That is what the seed=42 argument does above: the same seed yields the same GPU string, canvas hash, audio context and font set on every run, so a page that renders differently between two replays is telling you about your own code rather than about a random fingerprint draw. Pinning fingerprint fields covers forcing individual values while leaving the rest seed-derived.

Pair the frozen network with a frozen identity and a failing replay is reproducible down to the pixel. That combination is the whole reason to record a HAR in the first place: not to go faster, but to make a flaky test stop lying to you.

The caveat that decides where this belongs

A replayed HAR is a stale fixture. It serves the bodies that were captured, at the moment they were captured, and it never touches the live site again. That has two consequences worth stating plainly.

First, replay does not help you against a live server. Nothing in a route_from_har run reaches the internet, so it cannot pass a real check, defeat a challenge, or make a session look human to anyone. It is a tool for testing your own code offline, full stop. This project is built to look like a genuine Firefox driven by a real person, and that is why it clears most fingerprint, TLS and driver-layer checks: the engine is real, not patched over in JavaScript. But looking real is a live-session property. It has nothing to do with replaying a file, and replaying a file gives you none of it.

Second, the HAR itself is sensitive. It was recorded through whatever exit your session used, so it contains real, IP-tied responses: cookies, tokens, session identifiers, personalised bodies, and headers that name your egress. record_har_mode and content settings change how much lands in the file, but a full capture with embedded content is a complete record of a real session. Treat capture.har the way you would treat a credential file: keep it out of the repo, scrub it before it is shared, and do not commit one you have not read.

What HAR replay does not fix

The record-and-replay loop is orthogonal to the reasons a live session gets blocked. It does not improve IP reputation, it does not extend a per-account quota, it does not change your request pacing, and it does not alter behaviour or timing. Those are supplied by you: a clean exit and human-shaped interaction. If a live run is being detected, a HAR fixture is the wrong tool; work the detection checklist instead, and if a clean fingerprint still gets blocked, there is a page for that.

Where HAR replay earns its place is the test suite. A regression test that must run in CI with no network, a golden-file comparison of how your parser handles a fixed response, a reproduction attached to a bug report: all of these want a frozen capture and none of them want the live site. Keeping that boundary clear is the difference between a fixture and a false sense of security.

Conclusion

record_har_path captures a session to a self-contained JSON log, and route_from_har serves it back offline so a test is deterministic and network-free. Both work as-is on the patched Firefox this project ships, driven by stock Playwright. Pin the seed alongside the file and a failing replay is fully reproducible. Just remember what the capture is: a stale fixture that cannot face a live site, and a file full of real session data that has to be handled like one.

Short answers to the questions that lead here

Does replaying a HAR make my automation look more human? No. Replay never touches the live network, so it cannot influence any live check. Looking real is a property of a real live session, which is what the engine provides separately.

Can I use a HAR to bypass a check on a real site? No, and it is the wrong mental model. A replayed request is answered from your file; the real server never sees it. Use it for offline tests, not against live endpoints.

Is it safe to commit the .har file? Treat it as sensitive. A full capture holds cookies, tokens and IP-tied responses from a real session. Read it, scrub it, and keep it out of public repos.

Why is my capture.har empty or missing? The file is written when the context closes. If the process crashed or you never called context.close(), nothing was flushed.

How do I make unmatched requests fail instead of hitting the network? Pass not_found="abort" to route_from_har. The default fallback silently lets unrecorded requests reach the live site, which hides gaps in your fixture.

Does HAR replay speed up or fix a blocked scraper? No. It has no effect on IP reputation, quotas, rate limits or pacing. Those need a clean exit and human timing, which you supply.

Sources

See also: Configuration for proxies and the egress-derived timezone that shaped whatever you captured, the detection checklist for a live block a fixture cannot help with, and how to test bot detection without a false pass for asserting the right signal rather than the absence of a wrong one.

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. A HAR makes a test honest; it does not make a session real.

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.