The case for making your own apps

Casey Newton · rss · 2026-07-22 source ↗ Reader 🖥️

AI Bubble Debate AI Labor Disruption Malleable Software Near-Zero-Cost Code Non-Engineers Building Vibe Coding

As AI collapses the cost of building software toward zero, a large and growing share of the software people use — within a few years, plausibly a third to half of it — will be hyper-personalized, disposable software they (or their friends and colleagues) make themselves, which will hollow out commercial SaaS products whose value was mainly a mediocre one-size-fits-all UI while sparing those that hold irreplaceable data or can operate headlessly.

practical · mixed — An edited Q&A interview (dialogue/narrative framing) built around a sustained argument that people should — and increasingly will — build their own software; Newton's framing and Mann's answers are organized around a prescriptive/predictive thesis, not neutral description.

Structure

Part	Role
Editorial framing (Newton's intro): Kuyda's Wabi prediction that subscription apps disappear, tested here on the Mac desktop with Raycast/Glaze's Thomas Paul Mann.	context
Origin story: Raycast's 3,000+ open-source extension ecosystem as the pre-AI proof that 'giving people a tool to create their own tools changes the equation.'	evidence / precedent
Mann's diagnosis: commercial software is a 'compromise' — growth pressure forces feature bloat until apps 'get more in the way than they actually help.'	core argument / diagnosis
Concrete Glaze examples: a personalized to-do app, a support-ticket tool built on internal APIs, an internal-presentations app, scaling from individual to team use.	evidence
Market-size argument: app stores already hold 'millions of applications'; the YouTube analogy (creation democratized beyond what looked like a small, silly market).	supporting analogy / inference
The SaaS apocalypse discussion: which categories of SaaS survive (records-holding, or able to go 'headless') versus struggle (weak-but-necessary UI, in-between state).	qualification / refinement of the thesis
Jobs discussion: AI benefits the most junior and most senior engineers most; the middle is in a 'weird spot'; the engineer-vs-builder identity anecdote.	related but distinct argument, tests the thesis against a labor-market objection
Closing prediction and call to action: ask your AI what app you should build tonight.	conclusion / practical restatement

Key terms

disposable software Malleable Software
Cheap, often single-purpose or single-use software built for a market of one (or a small team), with no expectation of durability, polish, or a market beyond the maker — 'weird,' 'very personal' apps that are fine to throw away.
hyper-personalized software
Software built for 'a market of exactly one' — taking the good parts of many existing options and mixing them into something that fits one specific person's workflow, rather than the broadest common denominator a commercial product must serve.
headless (applied to a SaaS product)
A software service whose user-facing interface is bypassed — consumed instead through an agent or a personalized Glaze dashboard — while its backend/data/service logic continues to be used. Mann's diagnostic category for which SaaS survives disruption.
compromise (software as)
Mann's diagnosis that commercial software design is shaped by the need to serve a broad market and grow revenue, producing feature accumulation that serves the company's growth incentive more than any individual user's actual needs — 'a little tool suddenly becomes this glorious massive app that wants to do everything.'
everybody is a creator now
Mann's claim (echoed by Newton) that software-making is undergoing the same democratization video underwent with YouTube: what looked like a small, novelty market (amateur creation) turns out to dwarf the professionally-produced market once the skill/cost barrier drops.

Propositions

explicitWithin roughly three years, 30 to 50 percent of the software on a person's Mac could be software they made themselves.conf: high
extendsVibe coding has escaped the terminal · p5 — Mann's quantified 30-50%-of-your-Mac prediction extends Newton's own earlier observation (from his first month with Glaze) that collapsing build cost/skill makes previously not-worth-pursuing software newly viable — turning a personal anecdote into a market-scale forecast.
explicitCommercial software is fundamentally a compromise: companies chasing growth pile features onto products until the products actively get in the way of the specific tasks any given user actually wants to do.conf: high
explicitGiving people a tool to create their own tools qualitatively changes what becomes possible, as demonstrated pre-AI by Raycast's extension ecosystem growing to more than 3,000 open-source, community-built extensions.conf: high
implicitThe market for personalized software will end up dramatically larger than today's app-store market, following the same pattern by which video creation (YouTube) grew from a seemingly small, silly niche into a mass market that dwarfed professionally-produced content.conf: medium
explicitSaaS products that hold a company's or person's core records (email, documents, knowledge) will keep their value for a long time, while products whose UI can't be fully removed but also isn't strong enough to be a real differentiator are the ones most at risk of losing users to agents and personalized apps.conf: high
refinesWhy cultivating agency matters more than cultivating skills in the AI era | Max Schoening (Head of Product, Notion) · p11 — Where Schoening argues the SaaS apocalypse is broadly exaggerated because 'as a service' retains value, Mann's headless/records-holding-versus-weak-UI distinction refines that into a testable framework for which SaaS categories actually survive versus struggle.
explicitAI benefits the most junior and the most senior engineers the most, leaving mid-level engineers in a comparatively weak spot: juniors gain an outsized uplift from access to previously untapped intelligence, and seniors gain a 'farm' of agent-juniors they can steer in parallel.conf: high
supportsFragments: February 13 · p4 — Mann's independent claim that AI benefits the very junior and very senior most, leaving the middle in a 'weird spot,' converges with Fowler's mid-level-hardest-hit observation, adding Mann's specific mechanism (juniors: untapped-intelligence uplift; seniors: an agent 'farm' to steer).
explicitAn entire product category (e.g. to-do apps) can persist as unsatisfying for years — one person can try 200 different to-do apps and still be dissatisfied — because every existing commercial option embodies a different compromise, and personalization is now cheap enough to individually resolve that.conf: high
implicitThe boundary between 'an app' and 'a feature added to an app' is dissolving: under Glaze, adding new functionality to one's own app (e.g. a new kind of slide, a new presentation) is functionally the same act as building a new app, undercutting the assumption that software ships as a fixed, pre-defined set of features.conf: medium
explicitLarge, broad AI companies (OpenAI, Anthropic) are unlikely to dominate the personalized-app-making layer themselves, because a company serving everyone at once cannot go deep enough into any one product category to out-execute a focused, vertically-committed startup.conf: high
implicitNon-technical people discovering they can build software surfaces product ideas that professional software development never would have produced, because the idea only becomes conceivable once building it feels possible.conf: medium
explicitWhether a SaaS subscription actually gets canceled depends less on an abstract 'AI could replace this' judgment than on the practical fact that the tool has quietly stopped being opened — and even then, vendor upsell pressure (pitching AI features) can outpace actual cancellation.conf: high

Arguments

Within a few years, roughly a third to half of the software an individual uses will be software they made themselves.

Explicit premises:

Giving people a tool to build tools changes the equation, as shown by Raycast's extension ecosystem reaching 3,000+ community extensions even before AI made building dramatically easier.
App stores already contain millions of applications, showing enormous latent demand for narrow, specific tools once publishing is cheap.
AI (via Claude Code and Codex, integrated into Glaze) has collapsed the skill and cost barrier to building a working app to near zero.

Implicit premises:

People's software needs are heterogeneous enough that no single commercially-built product can satisfy most individuals' exact needs — mass-market software is structurally a compromise for almost everyone.
Given a real choice, users prefer software that fits them exactly over software that fits 'most people reasonably well,' once the cost of the former approaches the cost of the latter.
The rate of AI model and tooling improvement will continue rather than plateau before this transition completes.

Inferences:

If the pattern holds, software companies will need to compete on what users can't easily build themselves (data custody, integration breadth, headless infrastructure) rather than on UI/UX.
Traditional app-store curation and review models may not scale to a world where personal, one-off apps proliferate at this rate.

SaaS companies survive this shift if they either hold irreplaceable core records or can become 'headless' (consumed via API/agent without needing to win on UI); they struggle if their UI is a weak differentiator that nonetheless can't be fully removed.

Explicit premises:

Mann's own team has stopped opening many SaaS tools directly, instead consuming the same underlying services through agents or Glaze-built dashboards.
Data- and records-holding software (email, documents) remains necessary because rebuilding that infrastructure yourself at every company isn't worthwhile.

Implicit premises:

A vendor's UI investment, not its backend functionality, is usually what justifies charging a premium over accessing the same data/service directly.
Vendors that don't proactively offer a headless/API-consumption mode risk having users route around their UI via agents anyway, capturing none of the resulting value.

Inferences:

SaaS pricing built around seat-based UI usage may need to shift toward usage- or API-based pricing as more consumption becomes headless.

AI creates a structurally weaker position specifically for mid-level engineers, benefiting the very junior and very senior most.

Explicit premises:

Junior engineers get 'a massive uplift' from access to previously untapped intelligence and, if eager, can 'level up really quickly.'
Senior engineers get 'a farm of more junior people, in the form of agents' and can 'steer a lot more things in parallel.'

Implicit premises:

Mid-level engineers' comparative advantage before AI was execution speed and experience — a position that depends on neither the steep learning-curve runway juniors have nor the multi-agent delegation leverage seniors have, so AI erodes exactly what made them valuable.

Inferences:

If generalizable, this predicts a 'hollowing of the middle' in engineering career ladders, echoing labor-market hollowing patterns documented elsewhere.

Narratives

Newton's own Nightwing to-do app, built in Glaze, generates an image of the superhero completing each task the user checks off.
→ Illustrates concretely that the value of hyper-personal software can be pure delight and personal fit rather than objectively superior functionality — grounding the abstract 'disposable software' thesis in a specific, admittedly frivolous example.
Another Glaze user built an app that watches them via webcam and alerts them when they have bad posture.
→ Extends the same point further into idiosyncrasy: shows the tail of genuinely single-purpose utility software that would never justify commercial development but is trivially worth building for oneself.
David Pierce (Platformer colleague) tried roughly 200 different to-do apps and found each one had 'something good and something bad,' leaving him flip-flopping between them.
→ Supports the 'software is a compromise' proposition with a concrete, extreme case of category-level dissatisfaction that personalization is positioned to resolve.
A senior engineer on Mann's team who loved writing code had to make a deliberate decision to use AI more (and write code less), initially experiencing it as a loss of identity, before eventually finding joy in the new mode and becoming one of the team's heaviest AI users.
→ Illustrates the identity cost of AI adoption for engineers who derive meaning from the craft of writing code itself, and that adaptation is possible but requires an effortful 'mental shift,' not automatic enthusiasm.
Raycast's support team built its entire internal tooling on Glaze, pulling in APIs from Raycast's own services plus the internal knowledge base and manual, with AI drafting replies.
→ Shows the personalization pattern scaling from individual use to team-level internal tooling, supporting the claim that 'solve your own problem' software naturally scales because a colleague usually has the same problem.

Questions addressed

✅ What is the ceiling on how much personally-made software will replace professionally-made software?
Mann gives a concrete but explicitly speculative estimate: 30-50% of an individual's software within roughly three years.
✅ Does the SaaS industry face an existential threat from personalized and agent-made software?
Partially and conditionally: Mann differentiates rather than answering yes/no — SaaS that holds core records or can go headless persists; SaaS with a weak-but-necessary UI struggles.
✅ Will the AI model vendors (OpenAI, Anthropic) themselves crush platforms like Glaze by shipping a 'make me an app' feature?
Mann argues no: broad platforms serving everyone can't go deep enough into any one vertical to out-execute a focused startup building followership around one thing.
✅ What happens to engineering headcount as building software becomes radically cheaper?
Resolved only for Mann's own company: headcount has stayed flat while output has risen sharply — explicitly offered as one data point, not a general industry claim.
⭕ If cheap, hyper-personalized AI-made software becomes the default, what happens to the professional software industry's incentive to invest in genuinely hard, differentiated engineering — security, reliability, compliance — that personal apps typically lack?

Gaps

incompleteNo engagement with the maintenance burden of personal software: who fixes a self-made app when the underlying AI model changes behavior, an integrated API changes, or a bug surfaces months later — especially for the large population of non-technical creators Glaze is explicitly targeting, who by definition cannot debug it themselves.
Read next: Newton's own earlier hands-on Glaze review (extract-platformer-vibe-coding-escaped-terminal) and Fowler's vibe-coding bliki entry, both of which raise exactly this accountability/maintenance gap for AI-generated software more generally.
incompleteThe 30-50% figure and the SaaS-apocalypse framework rest entirely on Mann's own vantage point as Glaze's founder and CEO — a party financially interested in both claims being true — with no independent usage/retention data, churn statistics, or a non-Raycast data point offered anywhere in the interview.
Read next: Look for independent survey or analyst data on personal/vibe-coded app retention, actual SaaS cancellation rates attributable to AI substitution, and non-founder practitioner accounts.
incompleteSecurity and data handling of hyper-personalized apps built by non-programmers is never discussed, despite Mann's own admission that Glaze apps connect to internal knowledge bases, support data, and third-party APIs — precisely the combination of private-data access plus external connectivity the corpus's 'lethal trifecta' framework flags as risky, especially when built by people with no security training.
Read next: extract-fowler-bliki-vibe-coding#g4, which raises the identical gap (non-programmers being least equipped to recognize security trifectas) for vibe coding generally.
incompleteNo discussion of whether hyper-personalized software eventually re-converges on the same feature bloat it's positioned as an antidote to — especially once Glaze's own remix/sharing features (taking someone else's app and adjusting it) let personal apps spread and accumulate features the way commercial SaaS did.
Read next: Corpus material on Jevons paradox and feature-bloat cycles in software.

The case for making your own apps

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents