Browser Use is a real, funded company selling real products. AlohaJet is a two-week-old website describing a product you cannot download, install, or look at the code for. Everything below is a number you can check yourself.
AlohaJet says it scores 88% on WebArena, a standard test for web agents. The problem: that's higher than any real system has ever scored — and higher than actual humans doing the same tasks.
Higher is better. Every bar except the first is a published, independently listed result.
In plain terms: they claim to beat every expert team in the world, and
also beat humans, using cheaper and weaker AI models than anyone else —
and they published no method, no code and no task list. Their own web
page even has a hidden note in it reading
TODO: confirm benchmark numbers are current.
There's a likely innocent explanation. Their tool remembers and replays routes it has walked before. They ran each test task ten times on a website that never changes. If the memory isn't wiped between runs, runs two to ten are replays, not fresh attempts — which would inflate the score exactly like this.
We generated eleven possible advantages and then tried hard to destroy each one. Eight died completely. These three survived — but read the catch on each.
Browser Use is a US company and offers no way to choose where your data is processed — the word “region” doesn't appear once in their entire product documentation. Aloha is a Cyprus company. For European governments, banks and hospitals, that's a box Browser Use literally cannot tick.
The catch: being European isn't a product. There's no security page, no certifications, no data agreement — none of the things such a buyer actually asks for.
Companies already pay a lot for a locked-down browser installed on staff laptops — one rival, Island, is worth $4.8B doing exactly that. AlohaJet is a browser company. Browser Use sells to developers and has nothing in this market at all.
The catch: that's a different company. It needs device management, admin controls and audit logs — none of which exist yet.
In August 2026 a US appeals court sided with Perplexity against Amazon, reasoning that the user's own browser did the visiting, not the company's servers. If your agent runs on your laptop, the law currently looks at it more kindly.
The catch: this helps every local browser, not AlohaJet specifically — and Browser Use already gives away a tool that runs on your own machine too.
None of these are hard. All four are about being believable, and until they're done nothing else AlohaJet says can be taken seriously.
Put some code on GitHub — or stop saying “open source”
The page they link to as proof doesn't exist. Anyone can check this in ten seconds, and once they do, every other claim looks shaky. Takes half an hour to fix.
Make the download button work, or turn it into a waiting list
Right now the main button on the whole site leads to an error page. A waiting list that works is better than a download that doesn't.
Show how the 88% was measured, or take it down
Publish the tasks, the settings, and whether the memory was wiped between runs. A smaller, honest number would persuade far more people than a big unexplained one.
Lean hard on being European
It's the one thing a US competitor can't copy by writing code. But it needs the paperwork behind it — a security page, certifications, and a clear answer to “whose laws apply to my data”.
The whole market is shifting under both of them. This year OpenAI shut down its AI browser after nine months, and Google shut down its equivalent too. Both decided the feature belongs inside the browser and assistant people already use. Google is now building it straight into Chrome for roughly 200 million phones, and Claude does it inside Chrome as well — free.
Meanwhile the price of renting a cloud browser has fallen to about two cents an hour, and one company now charges nothing at all. So “we're cheaper” is not a strategy that survives.
What that means: AlohaJet's problem isn't that Browser Use is ahead. It's that the thing AlohaJet is describing is being given away by Google and Anthropic. The way out is not a better agent browser — it's picking a buyer those giants won't serve. That's the European, regulated, security-team buyer described above.
Browser Use is fast and reliable for one main reason: it never shows the AI a picture of the page. It turns the page into a short numbered list and lets the model point at a number. Everything else follows from that one decision. Below is the whole thing, in plain words — all of it checked against their actual source code.
An AI model can't see a website. Something has to translate a page into text the model can read. There are only three real options, and the choice determines your speed, your cost and your reliability.
The model has to read text out of an image and then guess pixel coordinates — “click at x=418, y=232”. It gets those wrong constantly, especially after the page shifts a little. This is how most computer-use agents work.
A real page is often one to two megabytes of markup. It won't fit in the context window, it costs a fortune, and about 99% of it is styling, tracking and wrapper divs the model doesn't need.
Strip the page down to only what you can actually interact with, give each one a number, and let the model say “click 12”. Cheap, fast, and there is no ambiguity about what got clicked. This is Browser Use's approach.
CDP stands for Chrome DevTools Protocol. Chrome has a built-in remote-control channel. When you press F12 and open developer tools, that panel isn't magic — it's a separate program talking to the browser over this channel. CDP is simply you speaking that same language yourself.
Mechanically it's a WebSocket connection to the browser where you send small JSON messages. “Give me the page structure.” “Dispatch a mouse click at this element.” “Tell me every network request.” Anything DevTools can do, your code can do, because it's the exact same interface.
The important part: this happens inside the browser, not on top of it. It is not a robot moving your physical mouse pointer. Nothing has to be drawn on a screen for it to work, which is why it runs fine on a server with no monitor attached, why it works on background tabs, and why it's fast — no rendering out to pixels and then interpreting those pixels back again.
Almost everyone in this space drives the browser through Playwright — Microsoft's browser-automation library. Playwright is excellent, but it's a layer built for testing: it waits for things, retries, auto-picks elements, and generally tries to be helpful. Each of those conveniences costs round-trips and hides detail an agent actually wants.
Browser Use dropped it. Their package has
no Playwright dependency at all — the only browser
library they ship is cdp-use, their own thin client that
talks CDP directly. One less layer between the decision and the
browser. This is the single most under-appreciated engineering choice
they've made, and it's checkable: it's right there in their
pyproject.toml.
It also explains the AlohaJet benchmark framing. Playwright MCP scoring 11% isn't a scandal — it's a tool layer with no planner attached. Browser Use isn't beating Playwright at automation; it replaced that layer entirely with something purpose-built for a model.
This is the core trick. They pull the page structure over CDP, throw away everything that isn't interactive, and hand the model something that looks like this.
[Current page: https://example.com/login] [1]<a>Home [2]<a>Pricing [11]<input type=email placeholder=Email address /> *[12]<input type=password placeholder=Password /> [13]<button>Sign in [14]<a>Forgot your password? ... and the model replies with: { "input_text": { "index": 11, "text": "me@example.com" } }
The number is a handle, not a guess
[13] refers to exactly one element. The model never has
to write a CSS selector, an XPath, or a pixel coordinate — the three
things language models are worst at and most likely to hallucinate.
It picks a number from a list it was just given. That is a far
easier job, so smaller and cheaper models can do it well.
The asterisk means “this is new”
A * in front of an element means it appeared since the
previous step. That's how the model notices a dropdown opened or an
error message showed up, without re-reading the whole
page. It's a small detail that saves a lot of tokens and a
lot of confusion.
Everything else is aggressively trimmed
Non-clickable, non-scrollable elements are dropped entirely. Text on
each element is capped at around 100 characters. Attributes are
filtered to a short whitelist. The insides of icons are collapsed to
<svg />. Every one of these is a deliberate
token-control decision, and together they're why a page that is two
megabytes of HTML becomes a couple of thousand tokens.
Hidden corners of the page still get found
Modern sites hide elements inside “shadow DOM” — sealed-off mini-documents that ordinary scripts can't see into — and inside iframes. Because CDP operates at the browser level rather than the page level, it can reach into both, and marks them so the model knows. This is a large part of why it works on real, messy sites rather than only on clean demos.
You give it a goal in ordinary language
“Find the cheapest direct flight to Lisbon next Friday and screenshot the price.” No script, no selectors.
It reads the page over CDP and builds the numbered list
Page structure plus accessibility information — the same data screen readers use, which is where the meaning lives: what's a button, what's a label, what's disabled.
The model picks one action
Click an index, type into an index, scroll, switch tab, extract data, or declare the task done. A short fixed menu, so there's little room to go off the rails.
The action is executed as a real browser event
A genuine mouse or keyboard event dispatched inside Chrome — not simulated JavaScript. Sites that check whether a click looks human are much harder to fool with fake events, so this matters.
Repeat until done
Re-read the page, mark what's new, ask again. Typical tasks run somewhere between a handful and a few dozen of these loops.
Reading the page is the easy half. The hard half is that websites actively try to block you. A plain automated browser is trivially detectable: it announces itself in dozens of small ways — missing graphics hardware, odd fonts, a giveaway flag that says “I am being controlled by automation”.
Their own published measurements make the scale of this clear. Against a set of protected sites, a standard invisible browser gets through 2% of the time. A normal visible one gets 50%. Their modified Chromium gets 81%. That gap is the actual product — and it is why they maintain their own Chromium build rather than patching from outside.
Add to that: home-internet IP addresses instead of data-centre ones (data-centre IPs are blocked on sight by many sites), automatic CAPTCHA solving, saved logins so the agent doesn't sign in from scratch every time, and machines that boot in under a second. That bundle is what the two cents an hour buys. The browser itself is nearly free; not getting blocked is the thing worth money.
beta import; the release notes state plainly that “the
existing Python agent remains unchanged.” The main package is still
essentially all Python — the Rust lives in a different repository.
One of these is a company with 111,750 GitHub stars, $17M in funding and two shipping commercial products. The other is a fourteen-day-old landing page where every single install path returns a 404.
AlohaJet cannot be better than Browser Use at anything today, because AlohaJet does not yet exist as software. The interesting question is not who wins — it is whether the position AlohaJet has staked out is worth building toward. It is. Just not for the reasons the site gives.
This report was commissioned to find angles where AlohaJet can beat Browser Use. Eleven candidates were generated and every one was handed to an adversarial reviewer instructed to kill it. None survived at full strength. Two survived in narrowed form, one survives only conditionally, and fourteen collapse outright. I also found that the best argument still available to AlohaJet appears nowhere on its website, while the three arguments it leads with are among its weakest.
The honest framing is this. Aloha Mobile Ltd is a real, competent,
ten-year-old bootstrapped company with 36 staff, roughly 10M monthly
active users, and — importantly —
a live Chromium fork rebased through Chromium 150.
That is serious engineering most agent-browser startups cannot do.
But AlohaJet the product is nine static HTML pages, four unstripped
TODO comments in production, two dead links, and a
benchmark number its own authors flagged as unconfirmed.
Browser Use, meanwhile, is not invulnerable, and this report does not treat it gently. It carries a critical CVE in its primary safety control with five further bypass reports still open. It has no named customers and no Series A seventeen months after seed. Its headline download figure fell by ~87% six days ago. And its own documentation contradicts its own marketing on the two numbers that matter most: the homepage sells 82% at 17¢ per task while the docs report 76.47% at $1.64. A real competitor could hurt it. AlohaJet is not currently a competitor.
/docs, no /security,
and no way to obtain the product.
Every claim below was checked directly over HTTP on 31 August 2026, and independently re-verified by a second agent. These are receipts, not impressions. Anyone can reproduce them in under a minute.
| Claim on alohajet.com | Check performed | Result |
|---|---|---|
| “Open source”, linked to github.com/alohajet | HTML fetch, api.github.com/orgs, /users, plus global repo search |
HTTP 404 — org never claimed. Repo search for “alohajet” returns total_count: 0. |
| “Download for macOS” — the site's primary CTA | Follow github.com/alohajet/releases/latest |
HTTP 404 — and the URL is malformed; it is missing the repository segment entirely. |
brew install alohajet, shown on three pages |
Homebrew cask + formula APIs; full index grep | HTTP 404 / 404 — no such cask or formula. The only Aloha entry in Homebrew is aloha-browser. |
| LLMdex: “one line: the base URL” | Site-wide search for a base URL, API reference, auth docs or repo | NOT OBTAINABLE — no endpoint, no docs, no code. A described concept with zero surface area. |
| Any release artifact at all | Grep every page for .dmg .pkg .zip .tar.gz; check npm and PyPI |
ZERO HITS — no binary anywhere. npm/alohajet and pypi/alohajet both 404. |
| 88% task success on WebArena | Read the page's own HTML source | SELF-FLAGGED — the page ships <!-- TODO: confirm benchmark numbers are current -->. |
| Pro $20 / Ultra $200 tiers | Read pricing page and its source comments | UNDEFINED — features read “Details coming soon”. Source notes prices “landed 2026-08-19”. |
| Enterprise: VPC, on-prem, air-gapped | Check /enterprise, /security, look for reference architecture |
MAILTO ONLY — /security returns 404. No SOC 2, ISO, threat model or SLA mentioned anywhere. |
| Documentation and developer surface | Request /docs /cli /changelog /llms.txt /blog /about |
ALL 404 — the sitemap contains exactly nine URLs. Subdomains are wildcard DNS serving nothing. |
| Public launch, press or community | Hacker News Algolia, web search, Product Hunt, Reddit, X, Wayback | NO FOOTPRINT — zero relevant hits. The Internet Archive has never captured the product site. |
Scroll table horizontally →
The company is not vapourware, and it would be lazy to imply otherwise.
github.com/AlohaBrowser/aloha-core is a live, verified fork
of Chromium — 98 stars, ~58 GB, default branch
abr-150.0.7871.46, with a rebase history spanning nine
major Chromium versions and a last push of 29 July 2026. Aloha Browser
holds 10M+ installs on each of two Play Store listings and a 4.7★ /
313K-rating iOS app. Aloha Mobile Ltd is a registered Cyprus company
(HE 352095), founded 2015, bootstrapped, ~36 staff. AlohaJet is hosted
on Aloha's own production infrastructure, so the attribution is
authentic — this is not a squat.
The capability is real. The product is not. Those two facts together are what make this worth writing about rather than dismissing.
Stripped of positioning language. Pros and cons are weighted by whether the thing can be used today, not by whether it is claimed.
A stated intention to build a local, open-source agent browser.
Positioned as “a browser for your agent” — headed or headless, CLI and MCP native, running on your own device with any model you choose. Its differentiator is LLMdex, a self-hosted OpenAI-compatible proxy that routes each step to the cheapest capable model and caches repeated tasks as replayable “routes”. Built by the Aloha Browser team.
STAGE — Pre-product. Site published ~18 Aug 2026. No code, no binary, no docs, no launch. StrengthsThe category-defining open-source framework, plus two shipping commercial products.
Sells Browser Use Agents (send a natural-language goal via API V4, receive completed work) and Browser Infrastructure (managed cloud Chromium with stealth, proxies, CAPTCHA handling, persistent profiles, live view and recordings). Around these sits an MIT-licensed Python framework, the Browser Harness coding-agent tool, SDKs and an MCP server.
STAGE — Shipping and commercial. Founded 2024, YC W25, $17M seed (Mar 2025), ~25 staff, San Francisco. StrengthsAlohaJet cells marked CLAIMED describe what the site asserts, not what was verified to work. Where a claim was tested and failed, it says so. The accent rail marks the stronger position.
| Dimension | AlohaJet | Browser Use | Note |
|---|---|---|---|
| Can you use it today | noEvery install path 404s. No binary exists. | yespip install browser-use, or a cloud dashboard and API key. |
The only dimension that matters until it changes. |
| Open source | claimedAsserted repeatedly. Zero published code; org unclaimed. | verifiedMIT, 111,750 stars, 347 contributors, 146 releases. | AlohaJet's claim is currently inverted against it. |
| Where page data goes | claimedStays on device; LLMdex self-hosted; models can be local. | bothCloud by default. The MIT framework runs fully local with Ollama. | Tie on architecture. Browser Use already offers the local path. |
| Logged-in session access | claimedRuns in your own browser, so your real profile and device identity. | partialProfile sync, storage-state export, TOTP, 1Password. Harness drives local Chrome. | Cookie sync moves the cookie, not the device's reputation. Real but narrow gap. |
| Legal standing for agent access | structuralOn-device is the fact pattern the Ninth Circuit accepted in Aug 2026. | weakerDatacenter servers initiating the connection is a harder argument to make. | AlohaJet's strongest asset — and it is nowhere on their site. |
| Anti-bot resistance | claimedA real residential machine with an aged profile is hardest to flag. | shippingStealth, residential proxies, CAPTCHA handling, fingerprint management. | Different strategies. Local wins per-session; cloud wins at scale. |
| Scale and concurrency | boundedBounded by one machine. No orchestration layer described. | managedManaged session lifecycle, concurrency tiers, regions, live view, recordings. | Local execution has a hard ceiling that is not a marketing problem. |
| Cost model | claimedFree tier, BYO keys, no browser-hour fee. Pro $20 / Ultra $200 undefined. | publishedMarketed at ~17¢ per solved task; own docs say $1.64 median real. $0.02 per browser hour. | Browser hours are commoditising toward zero industry-wide. Weak axis for either to compete on. |
| Model flexibility | claimedAny model; LLMdex routes to the cheapest capable one. | shippingArbitrary models including local via Ollama in the OSS framework. | Routing itself is a commodity: OpenRouter, LiteLLM, RouteLLM, Not Diamond. |
| Task replay / caching | claimedLLMdex “flywheel” caches routes and replays them, shared server-side. | shippingworkflow-use plus action caching. Browserbase Director does the same. |
Published as Agent Workflow Memory (ICML 2025). No invalidation policy is stated by AlohaJet. |
| Benchmark credibility | weak88% WebArena. No methodology, no harness, no task list. Self-flagged TODO. | partial82% on a self-defined BU Bench V1, but the harness is open-sourced. | Both market self-defined numbers. Only one publishes a reproducible harness. |
| Platform support | limitedmacOS only. Linux “on the runway”. | broadPython and TypeScript, anywhere. Cloud needs no local runtime at all. | macOS-only excludes servers and CI — the actual home of headless work. |
| Developer surface | noneNine static pages. No docs, changelog, llms.txt or API reference. | completeOpenAPI, SDKs, REST, CDP, MCP, webhooks, OAuth metadata. | Browser Use even publishes machine-readable context files for AI agents. |
| Security posture | absent/security 404s. No threat model, no certifications, no injection policy. | flawed but realSOC 2 Type II (window ended Oct 2025, now stale). A critical CVE and five open allowlist bypasses — published and trackable. | A documented weakness beats an undocumented one in any security review. |
| Company backing | bootstrappedAloha Mobile Ltd, Cyprus, 2015, ~36 staff, no funding, ~10M MAU. | fundedYC W25, $17M seed led by Felicis with Paul Graham, ~25 staff. | Genuine tie. Bootstrapped longevity vs venture speed — different risks, not worse ones. |
| Engine-level control | unusedMaintains aloha-core, rebased through Chromium 150 — with no agent work on it. |
in productionAlso maintains its own Chromium fork, and ships it as the stealth product. | A widely assumed AlohaJet advantage that does not exist. Browser Use got there and monetised it. |
| Data sovereignty | structuralCyprus parent, EU jurisdiction. Non-US controller by construction. | absentNo region parameter anywhere in V2, V3 or V4. US controller. No VPC or single-tenant SKU. | The one axis where Browser Use has nothing rather than something weak. |
Scroll table horizontally →
Each candidate advantage was generated by one strategist and then handed to an adversarial refuter instructed to kill it. Eleven were proposed. None survived at full strength. Two survived in narrowed form and one survives only conditionally. They are ordered by how much weight they can actually bear.
Browser Use's entire 47KB machine-readable product context contains zero occurrences of “region”, “residency” or “sovereign”. There is no region parameter in V2, V3 or V4, no VPC, single-tenant or air-gapped SKU, and the contracting entity is American. On every other axis Browser Use has a good answer or a partial one. Here it has none at all.
/security is a 404, there is no SOC 2, no ISO 27001, no DPA, no sub-processor list, no threat model. This angle is an open door that AlohaJet has not walked through.Enterprises already pay for a browser installed on the endpoint as a security control point. Island reached roughly $87M ARR and a $4.8B valuation with ~450 enterprise customers; Palo Alto paid ~$625M for Talon. Gartner expects secure-enterprise-browser adoption to rise from 10% to 25% of organisations by 2028. AlohaJet is a browser company with an engine team. Browser Use has no product in this category at all.
brew install on one developer's laptop. AlohaJet's entire current shape serves the wrong buyer.On 4 August 2026 the Ninth Circuit vacated Amazon's injunction against Perplexity's Comet, reasoning that it is the user — not the AI company — who accesses the protected computer, because “Perplexity's servers do not directly access Amazon's… the user's browser does the communicating.” On-device execution is the fact pattern that survived a CFAA challenge.
These are the arguments AlohaJet's website actually leads with. Each one was tested against what Browser Use already ships and against the available evidence. None of them survive.
Both companies market self-defined numbers. Both are audited here on the same standard.
“88% task success rate on WebArena”
This would simultaneously beat the published state of the art (~74.3%, WebTactix on DeepSeek v3.2, Feb 2026) by roughly 14 points and exceed the measured human baseline of 78.24% — while using materially weaker models than every system at the top of the leaderboard. The entire published methodology is one sentence. There is a specific, plausible and non-fraudulent mechanism that would produce exactly this artefact: LLMdex caches and replays routes, WebArena is a frozen self-hosted environment, and they ran ten repetitions per task. If the cache is not cleared between repetitions, runs two through ten are replays rather than independent trials. Nothing on the site says whether it was cleared.
“Open source” / “Free and open source”
Stated on the homepage, the download page and in the footer of every page. No code has been published under the AlohaJet name anywhere: the linked GitHub organisation has never been claimed, and a global repository search returns zero results. This is not “only some parts are open” — it is nothing.
Benchmark peer selection: Chrome DevTools MCP and Playwright MCP
Selective rather than accidental. IBM CUGA (61.7%) and OpenAI Operator (58.1%) publish numbers on this exact benchmark and are free to cite. Excluding every system that shares the benchmark while including two that have never run it is the signature of a chosen denominator. Against CUGA the story is 88 versus 62; against a tool server with no agent it is 88 versus 11.
“2.2× faster” and “54% fewer tokens”
Arithmetic on their own published table: 73 ÷ 39 = 1.87× against
Chrome DevTools MCP, and 84 ÷ 39 = 2.15× against Playwright MCP. The
headline “2.2×” rounds up the weaker of the two comparators. The 54%
token figure does check out — but only against Chrome DevTools MCP,
because the Playwright MCP token cell on their own table is an em
dash, never measured. Most damaging: the token-savings graphic is
labelled Illustrative comparison in the page source, on
a page whose headline is “Measured, not promised.”
“Hundreds of millions of users trust Aloha”
Conflates cumulative downloads with users. The company's own 2024 press release gives roughly 10M monthly active users against 250M+ lifetime downloads. The underlying distribution is genuinely impressive; the phrasing is not defensible.
“82% on Internal Bench Hard at 17¢ per solved task”
Contradicted by Browser Use's own documentation. Their Choosing an Agent docs page states the V4 agent at 76.47%, and — measured on the same model across the last 30 days of production — puts the typical customer's successful task at $1.64 on V4, $0.87 on V3 and $0.70 on V2. That is roughly a six-point accuracy gap and a tenfold cost gap between the marketed figure and the documented real-customer figure. “82% at 17¢” is best-case harness performance on the cheapest model, not what buyers experience. To their partial credit, the honest numbers are published — they are simply not the ones on the homepage.
V4 marketing copy describes V3 capabilities
Two headline features advertised for the current API are false for it, per Browser Use's own docs. “Structured JSON output with schema validation” — the V4 docs state plainly that V4 does not accept an output schema, and the V4 OpenAPI spec has no schema field; it returns a plain string. “Deterministic scripts that replay a learned workflow with no LLM cost” — the V4 Scripts documentation says replay “still starts an agent and uses tokens”. Both features are genuinely real in V3, which is the API customers are being steered away from. Notably, the zero-LLM replay that AlohaJet markets as its flywheel is something Browser Use built, shipped, and then regressed.
Implied traction from PyPI download volume
Downloads ran at 1.7–2.9M/day through 24 August 2026, then fell to 142K–352K/day from 25 August and stayed there. Control packages measured over the identical window show only normal reporting tail-off. The drop is a sustained, package-specific cliff, which indicates the headline figure was substantially one large automated consumer. The organic baseline is still large — roughly 5–10M/month — but anyone citing “50M downloads a month” is citing an artefact.
The allowed_domains safety boundary
CVE-2025-47241 (critical, May 2025) allowed the domain allowlist to
be bypassed via the userinfo portion of a URL, affecting a reported
1,500+ downstream projects. Fourteen months later, five further
bypass reports remain open covering data:,
blob:, javascript: and chrome://
schemes plus hostname boundary handling. Users should treat the
allowlist as advisory rather than as a security boundary.
Ordered by what would change the situation most per unit of effort. The P0 items are not product work — they are credibility repairs, and until they are done nothing else on this list can earn a hearing.
Claim the GitHub org and publish something — or delete the word “open source” 30 minutes
Register github.com/alohajet today and push whatever is real, even if that is only the CLI wrapper or the MCP server. If nothing can be published yet, remove every “open source” claim from the site until something can.
A 404 on the org you link to as proof of openness is the cheapest possible thing to fix and the most expensive thing to leave broken. It invalidates the trust argument that every other claim depends on.
Fix or remove the two dead install paths 1 day
Either publish a signed release artifact and a real Homebrew tap, or replace the “Download for macOS” button and the brew install snippets with an honest waitlist.
A primary CTA that 404s converts curiosity into distrust at the exact moment of highest intent. A waitlist that works outperforms a download button that does not.
Publish the benchmark harness, or pull the numbers 1–2 weeks
Release the task list, evaluator version, model snapshot dates, per-model breakdown, variance, cost and raw traces. Critically: state explicitly whether the LLMdex route cache was cleared between the ten repetitions per task.
88% is above the human baseline on this benchmark. Presented without methodology, it reads as either an evaluation error or marketing — and a sophisticated reader will independently reconstruct the cache-replay hypothesis, which is worse than saying nothing.
Strip the production TODO comments 1 hour
Four internal build comments ship in the live HTML, including TODO: confirm benchmark numbers are current on the benchmarks page and a personal Dropbox path in the privacy policy.
Any journalist, competitor or prospective enterprise buyer who views source finds the team's own doubt about its headline number, in writing.
Own sovereignty — the one gap Browser Use cannot close 1 quarter
Build the position around EU jurisdiction rather than privacy in the abstract: a Cyprus controller, EU-only processing, a published sub-processor list, a DPA, and a documented answer to “which legal system can compel this data”. Then start SOC 2 Type I and ISO 27001.
Browser Use's complete machine-readable product context contains zero occurrences of “region”, “residency” or “sovereign”, and there is no region parameter in any API version. This is the only axis where they have nothing rather than something partial, and a non-US parent is the one asset they cannot ship, buy or blog around in a quarter. Everything else on this list is a race they can win; this one they structurally cannot enter.
Publish a security page and a prompt-injection threat model 2 weeks
Document the blast radius of an agent operating inside a real authenticated profile, the isolation boundaries, what the agent may and may not touch, and how injected page content is contained.
Running in the user's live logged-in session is a larger security exposure than a sandboxed cloud browser, not a smaller one. Independent research has found every agentic browser tested vulnerable to indirect injection. Shipping the capability without the threat model is disqualifying in any enterprise review — and Browser Use's own CVE history shows this is where the category gets hit.
Give LLMdex a surface, and a cache invalidation policy 3–4 weeks
Ship a base URL, an API reference, auth documentation and a repository. Then publish how routes are validated, when they are invalidated, and what happens on a stale replay.
Every mature implementation of this pattern answers staleness with validate-then-fallback; Stagehand re-engages the model when a cached selector misses. The site describes only the happy path. A replay cache with no invalidation story is exactly what a sceptical reader assumes inflated the benchmark.
Re-benchmark against real peers 2–3 weeks
Add Browser Use, Skyvern, Stagehand, IBM CUGA and OpenAI Operator. Keep the MCP tool servers as a floor reference if useful, but do not present them as the comparison set.
The current peer selection reads as evasion to anyone who knows the space, which is precisely the audience being recruited. Publishing a narrower win against real agents would be far more persuasive than a wide win against un-scaffolded tool servers.
Define Pro and Ultra, or remove the prices 2 days
Both paid tiers list their features as “Details coming soon” beneath a firm monthly price.
A price without a feature list asks for commitment while offering none. Removing the tiers until they are defined costs nothing; leaving them signals that pricing was decided before the product.
Pick the enterprise buyer, and build for them specifically 2–3 quarters
If the enterprise tier is real, it needs MDM enrolment, a policy engine, DLP, audit logging and SSO — the Island-shaped product, sold to CISOs. If it is not real, remove it.
Enterprise “local” means the managed fleet or VPC, not brew install on a developer's laptop. Those are different products with different buyers and different security reviews. Serving both badly is the current trajectory.
Put the Chromium fork to work — you are behind here, not ahead 2+ quarters
Land agent primitives inside aloha-core: native accessibility extraction, deterministic action APIs, and engine-level containment of injected page content.
Correct a false internal assumption first. Browser Use also maintains its own Chromium fork and has already monetised it as the stealth product, clearing 81% of protected sites against 2% for stock headless. Owning a fork is not the differentiator; shipping something with it is. AlohaJet's fork carries the maintenance cost of a ~58 GB rebase and currently returns nothing — the last twenty org events are inbound forks and watches with zero pushes.
Ship Linux, and stop overstating the user base 1 quarter
Headless work lives on servers and in CI, which macOS-only excludes entirely. Separately, replace “hundreds of millions of users” with the real and still-impressive ~10M MAU figure.
The positioning is “headless when you want to ship”, but the only supported platform cannot run in the environments people ship to. And an inflated user claim invites scrutiny of every other number on the site.
Neither company competes in a vacuum, and the terrain shifted twice this year in ways that matter more than the head-to-head.
The browser hour is dead as a revenue line. Posted prices span $0.02 to $0.12 for an identical commodity — a 6× spread that only ever resolves downward. Kernel has eliminated idle-time and proxy charges; Notte prices sessions at literally $0.00 and monetises elsewhere. Meanwhile every vendor has converged on $10–12 per GB for residential proxies. Margin is migrating to proxies, credential and auth infrastructure, CAPTCHA handling and outcome pricing. Any strategy whose core claim is “cheaper browser execution” is competing for a number heading to zero.
The standalone consumer agent browser has already failed twice. OpenAI discontinued ChatGPT Atlas on 9 August 2026, a nine-month lifespan, folding its capabilities back into ChatGPT and Codex. Google shut down Project Mariner in May 2026, folding it into Gemini Agent and Chrome Auto Browse. Both giants concluded the same thing: the capability belongs inside the assistant and the browser you already own. Chrome Auto Browse now ships at OS level on Pixel 10 and Galaxy S26, heading toward roughly 200M devices; Claude in Chrome reached general availability on 26 August 2026 and takes autonomous actions using your existing logins.
That last point is the structural threat to both companies, and it is larger than either. The single most valuable thing AlohaJet claims — an agent operating inside your real, already-authenticated browser — is being shipped for free, at operating-system scale, by the two companies that own the browser and the assistant. Browser Use's defence is that it sells to developers building products rather than to end users, and that is a genuine defence. AlohaJet's current positioning does not have one.
But the local thesis has a proven enterprise form. Island reached a $4.8B valuation selling a browser installed on the endpoint as a security control point, and Gartner expects that category to grow from 10% to 25% of organisations by 2028. The counter-evidence is equally real: open-weight models fell from 19% to 11% of enterprise LLM usage between 2024 and 2025, the EU–US Data Privacy Framework defused much of the residency panic, and HIPAA now has a well-trodden cloud on-ramp. Buyers say local and frequently buy cloud. The durable version of the local argument is not privacy or cost — it is control, identity and, since August, legal standing.