Spokes.wiki Search About

Synthesis — Web Browsers

The evolving thesis. Spun out 2026-07-04 from the hub consumer-ai-assistants cluster — the browser-wars roundup that completed it, plus Gemini-in-Chrome. Founding corpus is small; the domain is large and volatile.

Current thesis — the browser's job is changing

For two decades a browser mattered because it was the on-ramp to search — the default search box was the prize, and "browser wars" meant fighting over that entry point. browser-wars-2026 argues that era is closing: as answers move to AI and tasks move to agents, the browser is becoming the surface where an AI acts on your behalf, and the contest is over whose AI that is. This is the browser-side face of the demand shift ../search-marketing-wiki tracks as ai-search-shift (discovery moving from links to model answers) — same tectonic move, seen from the client instead of the search-results page.

The axes the field sorts on

The roster, populated

Ingesting the full browser-wars-2026 roster (11 browsers) sharpens four things the founding thesis only sketched:

Open questions

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. The economics of a from-scratch engine. Two independent engines are now on record (ladybird funded, Falco hobby), both Rust, and the question has sharpened from "does anyone try" to "can one become daily-drivable, and at what cost." — needs: a funding or engineering report with headcount/timeline, or a standards-conformance measurement (T1/T2). A denominator arrived 2026-08-30, from the other end of the problem: mozilla-2024-financials prices the engine that already exists at $290.4M of software development in one year — a ceiling, since the line covers all Mozilla engineering. Nothing about a from-scratch build follows from it, but any funding report this edge eventually gets should be read against that number rather than against nothing. 1a. The other side of the search deal. New 2026-08-30, the residue of the quick win the financials half-closed. The ledger establishes the shape of Mozilla's dependency — one customer, 86% of contract revenue — and never names the customer, states no term, and says nothing about exclusivity or about what the US antitrust remedies against Google do to the arrangement. — needs: the court record from the search-distribution remedies, or reporting that quotes the agreement.

  2. Adoption evidence for either pole. Assistant-vs-agent is the spoke's organizing axis and it rests on launch coverage and pricing, not on usage. — needs: a T1/T2 usage or retention measurement for an AI browsing feature — anything beyond a vendor announcement. Sharpened 2026-08-16 by the six-page audit above: re-reading three vendor pages after six weeks produced one corrected detail and nothing else, so the volatile timer is not what is missing here. Every one of the six niche browsers has zero non-vendor sourcing, and brave — the largest of them — publishes no user number at all. The cheapest version of this edge is now: one independent figure for any single browser on the roster, which would also give the spoke its first T2 outside the incumbents. The cheap half CLOSED 2026-08-20 (research pass) by dap-browser-share (T1): opera 144,948 users and vivaldi 125 on a mandatory federal-government panel of forty million. The edge's own question is untouched — a share figure is not adoption of an AI browsing feature, and DAP names no AI product at all. Still needs: usage or retention for an assistant or agent feature specifically. The successor to the cheap half is edge 9 below.

  3. A number for a browser that hides on purpose. CLOSED 2026-08-21 (research pass) by homebrew-cask-analytics (T1) — a package-manager install count, which is one of the three instruments this edge named. Counting at brew install identifies the browser by the package being installed, so Brave's Chrome user-agent is irrelevant: 93,290 installs in the year to 2026-08-21, rank 52 of 22,411 casks, 1.8× Microsoft Edge on the same panel — an ordering no user-agent panel can produce. The two instruments disagree wildly below the incumbents (opera leads vivaldi 1,160:1 on dap-browser-share; Vivaldi leads Opera 1.9:1 on Homebrew), which is the useful caution: each is a channel, not a population. Original text: brave is absent from a 31-browser federal breakdown because it ships a Chrome user-agent by design, so every user-agent panel — DAP, StatCounter, Cloudflare, all of them — counts Brave as Chrome — needs: a measurement that does not read the user-agent. 9a. Brave's size outside one package manager. New 2026-08-21, successor to the above. What Homebrew supplies is a floor inside a macOS developer channel whose opt-out analytics the privacy-minded disproportionately disable, and most Brave installs arrive as a .dmg from brave.com. Nothing yet counts Brave on Windows, on Android, or in the general population. — needs: a Microsoft Store or Google Play install figure, a Linux distribution's popularity-contest count, or a behaviour-identifying panel study. HALF CLOSED 2026-08-22 (research pass) by brave-play-store-listing — one of the three instruments named, and it delivers less than the ask implied. Google Play publishes a bucket: 100M+ downloads (next rung 500M+, so a 5× range), 3.71M reviews, cumulative since launch, Android only, installs not people. Brave's own copy on the same page claims "over 100 million users," a vendor claim about a different quantity that happens to print the same digits — kept apart on the page. Still needed: a figure that is neither bucketed nor cumulative — a Microsoft Store count, a Linux popularity-contest count, or a behaviour-identifying panel. Nothing yet puts Brave in a population. HALF CLOSED AGAIN 2026-08-25 (research pass) by flathub-browser-installs (T1) — the Linux package count, the second of the three instruments named. brave 6,483,260 installs, 2.1× Microsoft Edge, against 1.8× on Homebrew: two package managers on two operating systems, sampling different people, agreeing on the ratio. That corroboration is what the Homebrew figure lacked, and it is the strongest thing this edge has produced. What it still is not: Flathub is one channel on a platform where most installs arrive as .deb, .rpm, Snap or Brave's own script, and the totals count events since listing, not people. Still needed, and now narrower: a Windows figure, or any count that is neither per-channel nor cumulative. The pair also got worse rather than better — opera leads vivaldi 1.5:1 here, having trailed 1.9:1 on Homebrew and led 1,160:1 on DAP. 9b. Three channels, three orderings, and no way to weight them. New 2026-08-25. Every non-user-agent instrument this spoke has found is a distribution channel, and the Opera/Vivaldi ranking now reverses across all three of them. Nothing tells us what fraction of any browser's users a channel represents, so the counts cannot be combined or reconciled. — needs: a source that states a channel's coverage of a browser's install base (a vendor's own channel breakdown, or a panel measuring how software was obtained).

  4. Whether the trust ceiling is payable. chrome-built-in-ai-dos-donts rules out on-device locality as the escape from prompt injection, so the cost has to be paid in gating, sandboxing or confirmation — and nothing here measures what that costs in usefulness. — needs: a T1/T2 evaluation of an agentic browser under adversarial pages.

  5. The incumbents' real response. CLOSED 2026-10-07 (research pass). Google's, Apple's and Microsoft's distribution moves are half the story and only partly sourced. — needs: first-party platform documentation or announcements (T3 acceptable — these are first-party facts). Hunted 2026-09-17 (as Most-wanted #31) — nothing ≥ bar. Firefox's attribution docs confirm the msstoresignedin key exists; the Public Data Report still publishes no channel breakdown. Not eligible again before 2026-10-01. safari-apple-intelligence (T3, MacRumors on Apple's WWDC announcement) and edge-copilot-integration (T1, the official Microsoft Edge Blog) close it. Successor, edge 4a: both sources are announcement-shaped — neither measures adoption of the features they describe, and Edge's 13.7% share figure credits distribution, not AI, leaving "does the AI feature move the number" unanswered for either incumbent. — needs: a usage or retention figure tied specifically to an incumbent's AI feature (not the browser generally).

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. Firefox and Gecko. CLOSED 2026-08-09 (research pass) — gecko from Mozilla's platform docs (T1) and statcounter-browser-share (T3) for the position it is in; firefox and mozilla written. Successor, narrower and unmet: who pays for the independent engine. Gecko's development has been funded largely by a search deal with the company shipping the dominant engine, and this spoke holds nothing on it. — needs: Mozilla's financial statements, or reporting on the Google agreement's current terms (T1/T2).
  2. The incumbents as browsers, not as AI features. CLOSED 2026-10-07 by the same two sources as edge 4. Chrome appears only through chrome-built-in-ai and gemini-in-chrome; Safari and Edge have no page at all. The "incumbents' real response" question cannot be answered against products the wiki does not hold. — needs: one page each, from release notes and platform documentation.
  3. Extensions as the contested surface. Ad blocking, agent add-ons and Manifest V3 all live in the extension layer, and no page names it. — needs: the WebExtensions/MV3 specification plus one write-up of what the migration cost blockers.
  4. A conformance scoreboard with a denominator. CLOSED 2026-08-11. interop-2026 (T1, the project's own README) and wpt-overview-bocoup (T2) supply both halves: the denominator is 56,552 tests / 1.8M subtests, so kitesurf's "215,000+" is about an eighth of the subtests and the page now says so. The method matters more than the number — Interop scores only tests that pass in every browser, so a single vendor cannot move it. Successor, unmet: nobody's engine but the incumbents' is on the board. Interop scores Chrome, Edge, Firefox and Safari desktop runs; ladybird and Falco appear nowhere in it, and each publishes its own progress in its own units. — needs: a WPT pass rate for an independent engine measured on the same runs, or the project's stated reason for excluding one.

The platform layer says the quiet parts

chrome-built-in-ai-dos-donts is the incumbent writing down the failure modes of in-browser AI as engineering requirements, and it lands on two of this spoke's open threads.

On-device does not buy trust. The strongest line in the doc is Chrome telling developers to "treat all LLM output as untrusted content" and to sanitize the full combined output, not just chunks, because "malicious code could be split across updates." This is a first-party model running on the user's own hardware, and it is still handled as an injection source — with streaming making it worse, since a payload split across chunk boundaries defeats a per-chunk filter. So the locality-equals-safety intuition that chrome-built-in-ai contributed to the privacy current has a hard limit: local execution protects your data from the server, and does nothing about the model's output attacking your page. That's directly relevant to the standing question of whether prompt-injection risk caps how agentic a browser can safely be — the answer isn't "run it locally."

Chrome is consistent across both layers. The UX rules — never replace UI without warning, never overwrite a draft without revert or compare, ship a stepper/history, "make the user the final editor," don't "force an AI-generated result as the only option" — are the assistant pole restated as build requirements. The same vendor markets the same restraint in gemini-in-chrome. Whatever one makes of the strategy, it isn't only positioning: it's written into the developer contract too.

One recommendation doesn't fit. The doc suggests "an artificial delay of one or two seconds if a response is nearly instant", because perceived effort reads as trustworthiness. On-device inference removes the round-trip and the advice is to restore some of it as theatre. It's minor and openly reasoned, but it's guidance to misrepresent work done in order to buy confidence, sitting beside rules about keeping the user in control. Recorded, not smoothed over.

A third role: the browser as trust intermediary

The spoke has sorted the field on assistant ↔ agent — does the browser answer you or act for you. email-verification-protocol doesn't fit either end, and reading it back against chrome-built-in-ai shows the earlier source didn't quite either. Both are the browser holding something and brokering with it: the model in one case, your first-party session with your email provider in the other. Paged as browser-as-trust-intermediary.

EVP verifies you own an address by having Chrome ask accounts.google.com with an ephemeral key and your existing cookies, then hand the site a signed token on form submit. The issuer never learns which site asked. The site learns nothing until submission. No code, no magic link, no round trip out of the flow.

Why this is a third role and not a variation. Assistant and agent features are sold to the user — perplexity-comet at $200/mo, opera-neon at $19.90. Intermediary features are given to the web, and the return is dependency: a site that retires its OTP flow needs a browser that implements EVP and a user whose provider participates. It's also the one capability a website can't route around by shipping better JavaScript, because the value is precisely that the browser sees both sides and the site sees one.

It complicates the privacy current. Privacy has been the challengers' differentiator here — brave's blocking, vivaldi's refusal of generative AI, ladybird's refusal of money. EVP's privacy properties are cryptographic rather than policy commitments, and they ship in Chrome. The current needs restating: what separates the privacy-first browsers is not that they alone provide privacy guarantees, but what they are willing to give up to provide them.

And it concentrates. The spec is open — WICG, an IETF draft, DNS-based issuer delegation so any provider can join. At launch Google is the browser, the email provider and the issuer, and Gmail addresses work with no configuration. "Open standard, one implementer" holds only as long as the implementer wants it to. Same shape as chrome-built-in-ai: a genuinely useful capability that also deepens the platform gravity this spoke tracks. Watch for a second issuer, or a non-Chromium engine shipping it.

The browser nobody opens

The spoke's thesis has been that the browser is becoming the surface where an AI acts on your behalf, and its axis measures how much the AI does unattended. kitesurf takes the premise one step further than the axis can express: Cloudflare shipped a browser with no user interface at all — no tabs, no chrome, no window — running on Workers, invoked per request through Browser Run, and tuned for context windows and token costs because a model reads its output.

Paged as agent-first-browser, a category beside the assistant↔agent axis rather than a point on it. What it changes in the thesis is the location of the acting: some of it happens in browsers nobody opens, bought by the request, where the buyer is a developer and the metrics are cost per page and throughput rather than trust and interface.

Two threads it touches without settling. Engine diversity gains its first corporately funded non-Chromium engine, for reasons that have nothing to do with winning users (see the open question above). And prompt injection is untouched: removing the human removes the last party who might have noticed a page instructing the agent reading it, which sharpens rather than eases the trust ceiling chrome-built-in-ai-dos-donts established.

Evidential standing: T3. A vendor announcement relayed by TechCrunch. The 3–7× efficiency claim is Cloudflare's, on Cloudflare's chosen workloads (screenshots and HTML extraction — the two tasks that most favour a cut-down engine), with no methodology published.

The one number in this domain that isn't self-reported

Nearly every figure this spoke holds was produced by the party it flatters. Cloudflare measured kitesurf. Google measured chrome-built-in-ai. Opera priced opera-neon. StatCounter's share figures ship with no methodology at all. The wiki has been recording that caveat page by page without having anything to contrast it against.

interop-2026 and wpt-overview-bocoup are the contrast. Apple, Google, Microsoft, Mozilla, Igalia and Bocoup run a shared test suite — 56,552 tests, 1.8 million subtests — in CI every few hours, with tests flowing both ways between the suite and each engine, and they score it on the tests that pass in every browser. A vendor cannot improve that number by shipping something first. The rule was chosen deliberately so that "everyone wins as interoperability increases" and the press cannot build a league table out of it.

Two things follow for this spoke. Conformance claims now have a denominator, which retroactively sizes kitesurf's "215,000+ tests" at roughly an eighth of the subtests. And the engine-diversity question gets a harder edge: the 2026 focus list is twenty ordinary platform features — dialogs, IndexedDB, scroll snap, WebRTC — on the list precisely because four mature, funded engines still disagree about them. ladybird and Falco are writing engines against that target, and neither appears on the board that measures it.

Worth noting what the shared metric costs: a score designed so that nobody loses cannot tell you which browser is better. It answers a question about the platform, not about the products this spoke spends most of its pages comparing.

The second one, and what it proves about the first (2026-08-20)

dap-browser-share is the other. The US General Services Administration publishes the browser breakdown of traffic to federal websites as open JSON with the query attached, and agency participation is required rather than opted into — forty million users, thirty-one browsers, no vendor anywhere in the chain. It hands this spoke the figure growth edge 2 asked for: opera at 144,948 users (0.36%) and vivaldi at 125 users (0.0003%), in a tail that also contains Waterfox at 22 and Maxthon at 33.

The more useful result is the one the dataset produces by failing. brave is not in it — not small, absent, from a list whose only exclusions are null-ish labels and counts under six. Brave ships a Chrome user-agent deliberately, because a distinctive one is a fingerprinting surface, so its users are inside the Chrome row and no user-agent-based panel can ever separate them. The audit of 2026-08-16 read Brave's missing numbers as a sourcing gap this spoke had not worked hard enough to close. It is not. The privacy pole of the roster is unmeasurable by the standard instruments as a consequence of what makes it the privacy pole, and the vendor's own MAU claim is the only figure that will ever exist for it.

That cuts both ways on the self-reporting complaint above. Where a browser keeps a distinct user-agent — opera, vivaldi, firefox — independent measurement is available and this spoke simply had not gone looking. Where a browser suppresses one, the absence of independent measurement is not negligence by anybody. Two different situations that had been filed as one.

The caveats are real and none of them is fatal. The panel is passports and tax forms rather than the web, which is why Edge sits at 14.52% and Internet Explorer at 1.88% in 2025; the feed froze on 2025-02-05 and has not moved since; and users are not the pageviews statcounter-browser-share counts, so the two must not be differenced.

Six pages, one afternoon (audited 2026-08-16)

The hub's standing backlog carried an item about short pages, and this spoke holds six of the corpus's twenty: sigmaos 129 words, zen 134, jatter 137, opera-air 139, brave 142, vivaldi 143. Audited as a group rather than as stubs, they turn out to be one artifact, not six defects.

All six were written on 2026-07-04, from browser-wars-2026 plus a first-party product page, and all six are tier: T3 and freshness: volatile. Nothing has been added to any of them since. So the spoke's thin-page problem and its floor problem (T3 13 against T1 4) are the same problem stated twice: a single day's ingest of vendor self-description, never revisited. None of them is padding — each names a real product, carries the thesis, and links out — and none of them can get longer without a source that is not the vendor.

Three were re-verified live this cycle, which is what volatile is a promise to do:

The finding worth keeping: six weeks of drift produced one wrong detail, one strengthened claim, and no decay anywhere else. For product pages of this kind the 60-day volatile window is doing very little work — what these pages need is not re-reading the same vendor page on a timer, it is one non-vendor source each. That is growth edge 2, and this audit is the clearest evidence for it the spoke has produced.

That finding is overturned, 2026-09-03. Kept above as written, because the reasoning was sound on the evidence it had and the correction is the interesting part. The whole 2026-07-04 cohort went past the volatile window on 2026-09-02, and re-reading all seven produced three corrections in one afternoon, one of them terminal:

  • openai-atlas no longer exists. It shut down on 9 August 2026 — three weeks before this check — and its agentic capabilities moved into a combined ChatGPT/Codex desktop app announced in March. It never shipped beyond macOS. The page described a live product for a month.
  • perplexity-comet's paywall came off. It was ingested as the roster's premium entry at $200/month; Comet went free in October 2025, with only unattended Background Assistants still behind the Max plan. The page had the economics exactly backwards.
  • opera-neon opened to the public on 2025-12-11, having been waitlist-only "Founders" access at ingest. Price unchanged at $19.90/month.

The remaining four (zen, jatter, opera-air, aside-browser) were unchanged, and aside-browser gained a tier list it had never carried.

What was actually wrong with the August reasoning. It generalised from a six-week sample of three pages to a claim about the window, and six weeks is inside the period where nothing had yet happened. Nine weeks caught a product death. The corrected rule is narrower and worth keeping: the volatile timer earns its cost on vendor pages, and the cost is one afternoon per cohort. Growth edge 2 is untouched — a non-vendor source is still what these pages need, and none of today's three corrections came from one.

The substantive finding underneath. OpenAI built the best-resourced agentic browser on this roster, ran it ten months, and folded the agent back into the assistant app. Perplexity kept its browser and moved the price line from capability to supervision — free to run an agent, $200 to run one unattended. Both are moves away from the browser-as-agent-container, from opposite directions, and the assistant ↔ agent axis this spoke is organised around now has its first evidence that the axis may be about who is watching rather than about what the software can do.

A fourth axis: the surface, not the user (2026-08-17)

Every category this spoke has built sorts browsers by who the user is or what the browser does for them — the assistant↔agent axis, the browser whose user is software, privacy subtraction, productivity rearrangement. carbonyl fits none of them, and not because it is marginal: it is Chromium, browsing ordinarily, with the output redirected into a terminal. No window server, no compositor, no GPU context; it runs over SSH and in a safe-mode console.

So there is a fourth axis, and the corpus had not been looking at it: where the rendering goes. It is the only capability in this spoke that no amount of money buys from any of the other twelve browsers, and it is orthogonal to everything the AI/privacy/productivity split measures.

It also inverts the engine-diversity argument. ladybird and falco-browser-engine are here because a three-engine web is fragile, and both answer by building an implementation from scratch at enormous cost. Carbonyl assumes the engine question is settled and spends its effort on the surface — it is Chromium precisely so WebGL and WebAssembly work immediately. Two scarcity claims, and the spoke had only carried the first: one says the scarce thing is an independent implementation of the web platform, the other says it is somewhere else to put the one we have.

And the maintenance caveat needs widening. Carbonyl has 19,426 stars and has not been pushed since 2024-07-01 — for a Chromium fork, where upstream ships security releases roughly fortnightly, that is a materially different fact than dormancy in ordinary software. The hub logged three abandoned high-star repositories on this one day, failing in three different ways, and the lesson is the same each time: stars measure attention at discovery. This spoke's standing volatile-data caveat covers pricing, features and geography; it should cover maintenance status too, since every browser page here is a snapshot of a product that may have stopped moving.

A fork that disagrees with its upstream about "secure" (2026-09-17)

titanium-browser is a one-maintainer Android browser: Vanadium's hardened Chromium, plus a patch layer whose main purpose is to restore extensions, Manifest V2 included, from any of three vendors' stores. Vanadium leaves extensions out on purpose — a smaller attack surface is the point of a hardened build — and Titanium's README concedes the ground by sending anyone who wants security to GrapheneOS. So the privacy-first row now holds both answers to the same question: brave ships blocking in the browser and keeps extensions; Vanadium removes them; Titanium takes Vanadium and puts them back. What the project does well is distribution — Play listing, attestable CI builds, in-place updates between the two — which is where a small fork's trust actually gets decided. 2.5k stars, no outside review, and the usual T3 footing.

Contradictions / tensions

Cross-spoke adjacency

Index — Web Browsers Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub consumer-ai-assistants cluster on 2026-07-04. Read synthesis first for the thesis (search→agency; the assistant↔agent axis; three currents — AI/agentic, privacy, productivity — that increasingly blur).

DefinedTerm (concepts)

SoftwareApplication (browsers)

Agent-first

AI / agentic

Privacy-first

Productivity / niche

Rehosted surface

Incumbent

Engine (shipping)

Engine (from-scratch; not a browser product)

Organization (makers)

Dataset (sources)

NewsArticle (sources)

TechArticle (sources)

Report (sources)

BlogPosting (sources)

Synthesis

Related pages

Links to

Linked from