Your tests stay Playwright, Selenium and pytest. We don't replace them. We make them enterprise-ready, and you own everything: code, evidence and history.
You own your code · You own your evidence
Write, record or import tests and run them on three browsers and real Android phones. Each run ends in one of 14 verdicts, and each failure keeps its report, logs and screenshot as evidence you can seal.
Optional: a mini PC of your own
USD 1,400 to 1,700 less in year one, then about USD 1,700 to 2,000 less each year. Estimate.
Windows 11 x64 or Ubuntu 24.04 LTS · 16 GB RAM minimum, 32 GB recommended
16 GB, about USD 340 to 385 · 32 GB, about USD 530 to 610
Estimate: a mini PC of about USD 350 plus about USD 60 a year of power, against one year of a cloud subscription with one parallel session. Your Nexus Studio plan, priced per machine and never per test, applies either way and is not in the figure; neither is your time looking after the box. Prices read on 2026-09-30: BrowserStack Automate, Desktop & Mobile, USD 175 a month billed annually, USD 2,100 a year, its entry option (browserstack.com/pricing). Sauce Labs Virtual Device Cloud, USD 149 a month billed annually (USD 199 month to month), 1 parallel test, USD 1,788 a year (saucelabs.com/pricing). Mini PC prices read on 2026-10-06, retail, x86-64 with a 500 GB SSD or more: about USD 340 to 385 with 16 GB of RAM, about USD 530 to 610 with 32 GB, the size we recommend. Power, USD 5 a month, an example.
Most test runners paint the same red for a slow network, a dead browser and a real defect. Someone loses a morning finding out which. Here a TIMEOUT is not proof the app is broken, a BLOCKED is the site defending itself, a CRASHED is the runner. Each one sends you somewhere different, so each one has its own name and its own color, the same ones the Studio paints.
78 runs from one real workspace, 2026-09-27 to 09-29: 20 runs.json files across chromium, firefox, webkit, Selenium, an Android phone and two API suites. Pass or fail would have painted eight of them the same red; here only six say the app broke. The line beside each name says its group without the color: solid for a pass, dashed for a pass with a reservation, double for a failure, dotted for no verdict.
A stopped, empty or broken test is never counted as a result, and a broken one tells you the exact error and line.
Evidence
Each failed run files a forensic report, its metadata, the telemetry and the raw log, plus a screenshot of the frame where it failed. A real failure keeps the video of the run as well: BREACHED, BREACHED_FLAKY and CHAOS_BREACHED by default, TIMEOUT and BLOCKED when you turn them on in Settings.
What one failure files
The real folder of the run replayed at the top of this page, as the Evidence Locker opens it: bug-reports/chromium/TC91 - Example more info/2026-09-30_13-26 [TIMEOUT].
Everything one failure leaves behind: the files, History and the housekeeping →
Chromium, Firefox and WebKit, plus real Android phones on USB with their screen mirrored on yours. Desktop + real Android device runs the PC and the phone at once and files the evidence per identity. iPhone and iPad run as emulated profiles on WebKit, not on physical Apple devices.
Every file of a record is hashed and chained to the one before. Verify names the first record that changed; with no ledger it says UNVERIFIED, never a silent OK.
A retry that passes is FLAKY, not SECURE. A chaos run where no fault landed is never called resilient.
A slow network, an offline blackout or a throttled CPU injected into the browser. A registered Docker container paused, stopped, killed or restarted while the recovery is measured. Load with latency percentiles and Web Vitals sampled on a real page under that load.
An autonomous scanner for targets you are authorized to test. SSRF, command injection and blind XSS count only when a callback reaches a listener on your own machine. Findings export as SARIF, each with a replay script.
The AI writer uses the AI command-line tool you already have (Claude Code, Codex or opencode), the one you are already signed into, so there is no API key to paste; your plan with that tool still applies. Or a model: local Ollama by default, Anthropic or OpenAI when you choose. A test it writes passes the Vet, can rehearse as a draft, and is saved only when you accept it.
Why Playwright and Selenium
Moving a suite into a proprietary low-code format is a risk many teams will not take. Nexus Studio keeps the code standard and builds the platform around it, on your machine.
Playwright, Selenium and pytest are open-source frameworks that many teams already use. Nexus Studio runs them. It does not replace them.
Your engineers and testers read a test from the Recorder the way they read their own: a page, a locator, an assertion. There is no new language and no proprietary test format.
You can hire for Playwright, Selenium or pytest, not for a vendor's tool. A new hire who knows the framework can read and change your tests without learning Nexus Studio first.
Each framework has public documentation, an issue tracker, and answers and libraries from its community. All of it applies to your tests as they are.
The team keeps its language, its locators and its habits. What changes is around the test: verdicts, evidence and scheduling. The Test Adapter brings in the tests you already have.
The Recorder and the AI writer produce ordinary test files in your workspace folder, and your team reviews them like any other code. The Vet, a deterministic check, reads AI-written code, and nothing is saved until you accept it.
Tests, evidence and history are files in your workspace. Python tests run under plain pytest as they are, and a Playwright JavaScript test runs without Nexus once its import points at @playwright/test. Selenium JavaScript tests need a few lines to start a driver, and API tests use the Studio's own format, so another runner would need them rewritten.
Nexus Studio@playwright/test and the file runs without Nexus.The engine that adds all of this runs on your machine, beside your tests. What the files look like, and how to leave
Maintenance
Selectors break, tests flake and evidence piles up. Each has its own tool in the Studio, built on your own history and your own files.
History ranks each test on each identity by how often it flips, breaks or costs time. A logistic regression over your own runs adds a risk score, with an EWMA fallback when data is thin, and says how sure it is: one run reads as low confidence.
Lists the weak and medium locators of the test you have open, one at a time, with the line and the reason, before they break. It reads the editor's buffer, so unsaved edits count.
When a step fails, stronger candidates are ready, strongest first and XPath last. Pick one and it replaces the selector in the editor in one click, as an unsaved edit; the earlier version stays in the Timeline. Or re-record from that line.
A pass keeps a screenshot and its metadata, 5 by default. A failure keeps its report, metadata, raw log and screenshot, and a real failure its video (you choose which verdicts in Settings), 20 runs per test by default. Older files go on their own, within the caps you set under Evidence & Retention, and all evidence together stays under one budget, 20 GB by default. Every default, on the evidence page
Estimate
We have not measured a saving on your suite, so we do not quote one. Enter how your team spends its time today and what you expect with the Studio, and the arithmetic is done below.
Example values, edit them. The minutes are our assumptions, not measurements.
By hand h a month
With the Studio h a month
Reduction%
An estimate from your numbers, not a measurement.
How it adds up: for each of the three kinds of work, how many a week times the minutes each, summed, times 52 weeks over 12 months, divided by 60. The reduction is one minus the Studio hours over the by-hand hours. It counts these three kinds of work only: not writing tests, and not waiting for runs to finish.
Pick one to see what the Studio does about it, and on which screen.
A failure that passes on retry is FLAKY, not a bug, and files nothing. With Jira or Azure Boards switched on, a failed run opens one issue per test, TIMEOUT included; a repeat failure comments on it and a clean pass closes it. BLOCKED, UNREACHABLE and CRASHED never do, and neither does a test that could not start, which is not a result at all.
Patrol, Evidence Locker
A 403, a 429 or a challenge page from Cloudflare, reCAPTCHA or hCaptcha reads BLOCKED: an environment verdict that never files a ticket. Automation signals are hidden by default.
Evidence Locker, Settings
When a step fails, stronger candidates are ready and Repair line swaps one in, strongest first and XPath last. Locator health flags weak ones before they break.
Patrol
Each failed run files its report, metadata, screenshot and raw log in one place, a real failure keeps its video too (you choose which verdicts in Settings), and a SHA-256 chain shows nothing changed afterwards.
Evidence Locker
The forensic report puts the failure, the telemetry and the page errors together, and the AI you choose can add an analysis. Real bug or False alarm takes one reason.
Evidence Locker, AI Providers
One Patrol run covers Chromium, Firefox and WebKit, one after another by default or up to three at once, plus real Android phones, with one verdict per engine.
Patrol, Devices & Mobile
Smart Monkey scans an authorized target for XSS, SQL injection, SSRF and more, confirms blind findings on a listener on your own machine, and exports SARIF.
Security
Chaos Lab slows the network, cuts it or throttles the CPU mid-test, and pauses or kills a Docker container while it measures the recovery.
Chaos Lab
Web Vitals are on by default and judged against budgets. Load adds p95 and p99 latency and samples a real page while the load runs.
Load, History
There is no proprietary format to migrate into or out of: the Recorder, the AI writer and the Test Adapter all produce Playwright, Selenium or pytest code.
Test Adapter, Nexus Recorder, AI writer
The Studio runs on your Windows 11 PC or mini PC. From it, the engine runs the suites on that machine, in Docker on it, or over SSH on another machine on your network or VPN. The license check sends a machine hash, a nonce, a token and the machine's host name, OS and Studio version, never test data.
Docker screen · Own your infrastructure
Six places a QA lead works in every day. Each frame is drawn from the Studio's own labels and a real workspace. The Studio tour covers every screen.
The Nexus Recorder writes Playwright or Selenium code as you click, with no AI in the loop. The AI writer uses the AI tool you already have, and the Test Adapter brings in the tests your team wrote. A deterministic review, the Vet, reads all three, and holds AI-written code to the strictest bar.
Record a test
Pick the target and the engines, and Patrol runs them one after another, or up to three at once. New tests rehearse in the sandbox first, where a run never counts in history.
Target
Desktop matrixMobile emulatedReal Android deviceDesktop + real Android device
headedvisuala11ysecurityno retryno stealth
Run patrol for real (1 TCs)Watch keeps re-running a selection on a timer and alerts on Telegram or Slack only when a state changes. It runs as its own process, so closing the Studio does not stop it. For pipelines, the CI/CD screen writes the GitHub Actions or GitLab CI file for you.
Watch
re-runs the selection · alerts only when a state changes
RepeatOnce, on a date
Each failure is filed where you can read it: the forensic report, the metadata, the telemetry, the raw log and, when your AI answers, its analysis: the settings the Studio ships with switch it on. Triage asks for a reason, and a Real bug or False alarm verdict joins the chain of custody.
| Test | Identity | Verdict | When (UTC) |
|---|---|---|---|
| TC98 | device-android-phone | TIMEOUT | 09-29 00:24 |
| TC03 | firefox | BREACHED | 09-27 19:06 |
TC03 · chromium · 2026-09-27_13-46 [BREACHED]
metadata.json 1.3 kB · report.md 3.0 kB · screenshot.png 46 kB · video.webm 469 kB
Real bugFalse alarmArchive…Chaos Lab injects faults into the browser, or pauses and kills a registered Docker container while it measures the recovery. Load reports percentiles and Web Vitals under load. Smart Monkey scans targets you are authorized to test.
Chaos Lab
History keeps each test on each identity, run after run. The flakiness predictor scores the risk from that same history, and says how sure it is. The Executive Report turns it into a release call: go, caution or no-go.
| Test | chromium | firefox | webkit | Android phone |
|---|---|---|---|---|
| TC02 | 9/9 SECURE | 7/8 SECURE | 6/6 SECURE | 1/1 SECURE |
| TC03 | 5/7 SECURE | 4/7 SECURE | 5/7 SECURE | no runs |
| TC04 | 6/6 SECURE | 6/6 SECURE | 6/6 SECURE | 0/1 TIMEOUT |
Nexus Recorder
Click through your app in a real browser, point at what must be true, and watch every selector get graded before it reaches the file. On every real run, the Studio keeps the screenshot and the verdict for you, and a failure keeps its logs too.
Pick a URL, desktop, emulated mobile or a real Android phone, and the kind of test to write. The test ID is reserved for you.
Record a test · Start recordingA real browser opens. Every click is written as code, live, in a panel next to it.
Recording · Undo · Save & closeArm an assertion, then click the element. The expected value is read from the page, never typed.
Assert visible · Assert text · Assert valueEach step shows how strong its selector is, how many elements it matches right now, and why. Fixed rules grade it, not an AI.
ROBUSTMEDIUMFRAGILE1 matchRecording saved offers Run in sandbox, headed. After three sandbox runs, which write nothing to evidence, you press the first real run. API recordings have no sandbox: they run for real on the first press.
Ready for their first real run · Run…Every real run keeps a screenshot, a classified verdict with an honest retry, and a History row. A failure adds the report and the console and network errors; a real failure keeps its video; you choose which verdicts in Settings. Nothing is captured while you record.
Evidence Locker · failure screenshot · session recordingRun patrol streams the console with a live status per test; tick headed to watch it. Locator health lists weak selectors, and Repair line swaps one in.
Run patrol · headed · Locator healthThe Studio runs on your Windows 11 PC or mini PC. Its engine runs the suites there, in Docker on that machine, or over SSH on another machine on your network or VPN. Tests, evidence and history stay on those disks.
Fourteen names separate a slow network, a bot challenge and a dead browser from a real defect, and Jira hears about failures and timeouts, never about a blocked site or a dead browser.
A retry that passes is FLAKY. A chaos run with no fault injected is NOT_INJECTED. Self-healing proposes a fix and leaves the run red until you choose.
Every file of a record is hashed and chained. Verify names the first record that changed, instead of asking you to trust a dashboard.
Your own AI tool or a local model writes. A deterministic Vet checks the test, a draft run can rehearse it, and nothing is saved until you accept it. The Recorder uses no AI at all.
Load, security scans and infrastructure chaos start only after an explicit authorization, and that authorization is recorded beside the evidence.
localhost
BASE_URL points the suite at your app on localhost or on a private staging URL, as it is. There is no tunnel to open and no staging URL to expose to a cloud QA vendor.
This machine
npm run devlocalhost · started before the run, stopped after itNEXUS_APP_START starts your app before every run and stops it with its whole process tree when the run ends, fails or is interrupted, Ctrl+C included.
The run waits until the ready URL answers, any status below 500, asked every 0.5 s. If the app cannot start, the run never starts: CONFIG_ERROR APP_DID_NOT_START, with its last 40 log lines.
NEXUS_APP_REUSE=true uses an app that already answers and leaves it running; false stops the run if something else answers there.
Each run records your app's package.json version and the git commit it tested.
In Docker the app still runs on this machine and the container reaches a localhost BASE_URL as host.docker.internal. Over SSH it starts on the remote host.
One program and its arguments, no shell: npm run dev, pnpm dev or node server.js. nexus app check, Test it in the Studio, starts it, waits and stops it. The whole log stays in evidence/run-context/.
On its own, Nexus Studio makes one connection, carrying zero test data: a license check. The key goes out once, to activate; each later check sends a machine hash, a nonce, a token and the machine's host name, OS and Studio version.
Getting started
From the trial to a scheduled run, in the order the Studio puts them. How long each step takes depends on your machine and your app.
Start the 30-day trial in your account, download the Windows installer and paste your key to activate. Then open a workspace folder: Setup installs its packages and browsers once.
Download for Windows · Activate · New folderOpen the Nexus Recorder, pick a URL and click through your app in a real browser. Every click is written as Playwright or Selenium code, with no AI in the loop.
Nexus Recorder · Record a test · Start recordingA new test starts in the sandbox, where a run writes nothing to evidence or history. After three rehearsals it is ready for its first real run. API recordings have no sandbox.
Run in sandbox · Waiting to leave the sandboxRun it on Chromium, Firefox and WebKit, and on a real Android phone if one is plugged in. A real run keeps a screenshot, a verdict and a History row.
Ready for their first real run · Run…Each engine gets one of 14 verdicts. A pass, a pass with a reservation, a broken application and an environment problem are told apart. A TIMEOUT is not proof the app is broken.
Patrol · SECURE · FLAKY · TIMEOUTA failed run is filed in the Evidence Locker with its report, metadata, telemetry, raw log and screenshot. Mark it Real bug or False alarm, and the verdict joins the chain of custody.
Evidence Locker · Report · Metadata · Telemetry · Raw LogWatch re-runs a selection on a timer and alerts only when a state changes, even with the Studio closed. For a pipeline, the CI/CD screen writes the GitHub Actions or GitLab CI file.
Watch · CI/CDIn plain words: Nexus Studio is a Windows app that runs the automated tests you already write in Playwright, Selenium or pytest, and tells you, in one of 14 clear verdicts, whether a failure is a real bug, a flaky test or just a slow network. It keeps proof of every failure and runs on your own computer, with a 30-day free trial and no card needed.
Individual is $79 a month for 1 machine, Team $199 a month for 3 and Business $599 a month for 10, with 2 months free when you pay for a year. Enterprise is quoted. No charge per user or per test, and a 30-day trial with no card.
Launch offer: once the 40 Founding Customer seats are filled, yearly plans get 30% off the first year, through December 31, 2026.
Before you commit: Support and longevity, or compare Nexus Studio with the other ways to test.
Free for 30 days, with every feature, on your machines. No card needed.
Ask us a question