The clone wars: what holds when AI rebuilds software

Early open-source alternatives to Adobe and Office show why feature replication deserves a fresh test. Production quality remains unproven; defensibility increasingly needs evidence beyond code.

MoatDefender / Published 2026-10-17 / Updated 2026-10-17

What the October headlines establish

Reports dated 11–14 October 2026 describe rapidly built, early open-source alternatives to Adobe and Microsoft Office applications. They do not establish that either mature suite has been fully reproduced. The title poses the strategic question raised by that building speed: if a capable rival can reproduce more of your visible product in weeks, which part of your advantage still takes something scarce to obtain?

TechSpot reported Brandon Thomas's ArtCraft and the Rust-based WordCraft, GridCraft and DeckCraft, built using Claude. WordCraft's repository describes its approach as a clean-room implementation using public specifications and observed behaviour. That is the project's own account of its process. This article makes no independent finding about that account or the legality of any implementation.

The Reverse Engineer Anything repository also connects coding agents to analysis tools including Ghidra, Hopper and IDA. That demonstrates an available tooling approach. Its presence does not establish that agents can reproduce every application, access private systems or obtain legal clearance to reuse what they discover. Each claimed capability still needs a bounded test.

The production gap remains material

WordCraft's roadmap calls the software alpha and gives a self-estimated readiness of about 60% toward replacing Word for real work. It names remaining work on real documents, file fidelity, pagination, stability and printing. That percentage is a developer estimate, rather than an independent benchmark. Feature coverage and professional reliability answer different questions; a familiar ribbon cannot establish both.

A buyer should test the files and workflows that matter: round-trip a difficult document, preserve its layout, validate calculations, check accessibility and investigate failures. Enterprise adoption also involves migration, procurement, support and distribution. The news challenges the cost of creating a credible starting point. It leaves the effort of earning dependable use open. Founders should measure that remaining effort before describing every feature or every line of code as a lasting barrier.

The specification paradox

Our inference from MoatDefender's research is the specification paradox: documentation, public interfaces and executable tests can help a rival understand what to reproduce and check whether it works. Cloudflare's February 2026 vinext account illustrates the mechanism. It reported implementing much of the Next.js API surface quickly, while explicitly calling the result experimental. The team identified documentation and reusable tests as enabling conditions.

This is a strategic inference from bounded cases, with uncertainty about how widely it applies. It gives founders a practical experiment: provide a permitted specification and ask a credible rival to reproduce the relevant outcome. Measure fidelity, human correction and operating cost. Good documentation still benefits customers and maintainers. The implication is to find the advantage that remains when documented features become easier to reproduce, then gather evidence for that advantage.

Cloudflare: vinext development account and experimental limits, 24 February 2026

Rights, licences and approvals

Software IP still matters. WIPO's guidance includes computer programs among copyright-protected works and distinguishes protected expression from ideas and methods. Faster implementation does not settle copyright, patent, contractual or trade-secret questions for a particular product. A clean-room label is no substitute for examining the facts with qualified counsel. The strategic assessment and the legal assessment need their own evidence.

Access also has conditions code cannot grant itself. Ask which data rights, licences, customer permissions and operating approvals the offer requires, who holds them and when they expire. The FDA's list of AI-enabled devices illustrates that marketing authorisation is a distinct regulatory fact. It does not make every approval a moat. A company must show that its particular access benefits customers and imposes a meaningful constraint on otherwise capable rivals.

Atoms, energy and audited track records

An implementation cannot create an available industrial site, commissioned equipment or a working grid connection. The IEA's Energy and AI report describes long connection queues and infrastructure lead times. Those constraints explain why physical access deserves examination. They establish no automatic advantage for a company: ownership, usable capacity, contractual conditions and alternative locations all affect whether the resource creates a barrier.

Audited track records deserve the same discipline. Dierickx and Cool's asset-accumulation theory examines assets built over time and the difficulty of compressing that process. Applied here, our hypothesis is that relevant operating evidence can matter when customers require it. Inspect the observation period, failures, independent review and scope. A rival can copy a report's appearance; earning an equivalent record may require real operation. Years elapsed alone prove little.

Liability absorption and learning loops

A service can become more valuable when someone credible takes responsibility for failure. Armilla describes AI liability insurance and performance warranties, with coverage subject to terms and exclusions. These products demonstrate that risk transfer can surround software. They do not prove that an insured vendor has a moat. Examine who owes the remedy, the available financial backing, exclusions and whether the arrangement actually changes customer choice.

Learning loops offer another candidate. Cursor describes improving its Tab system using feedback from accepted and rejected suggestions. That is a company report about a particular process, rather than proof that every dataset compounds. Ask whether the company may use the feedback, whether measured outcomes improve and whether competitors can obtain equivalent signals. The potential barrier is a useful, controlled learning process whose results persist against a strong baseline.

Identity-bound networks and complementary assets

Replicating a network's interface does not automatically recruit its counterparties or obtain their authority to transact. Google's Agent Payments Protocol announcement explicitly addresses authorisation, authenticity and accountability. It illustrates the importance of verifiable mandates around agent actions. An open protocol itself offers no exclusive position. A founder needs evidence that valuable relationships, permissions or settlement roles remain difficult to reproduce as competitors adopt the same standard.

Complementary assets connect these possibilities. Teece's research explains how value from innovation can accrue to owners of assets needed to commercialise it when imitation is easy. For a software company, the candidates might include distribution, service capacity or an installed operating network. Test their control and relevance. A promised channel, an unsigned partnership and a functioning distribution relationship support different claims. Every proposed moat needs a customer benefit and a specific obstacle to neutralising it.

Find → Design → Improve → Protect

Begin with Find: identify the benefit and the evidence behind your current advantage. Design: choose a defensible mechanism around the outcome customers need. Improve: test the weakest claim and record the result, including counterevidence. Protect: maintain the rights, relationships, operating capability and accountability that support it. This is an agenda for work, with no guarantee of durable advantage.

Use the Moat Index to structure that examination beyond code, alongside its documented principles and limits. A reading is an evidence-based assessment of defensibility, with uncertainty. It cannot turn alpha software into a production replacement, resolve an infringement dispute or predict investment returns. The useful question after the clone headlines is concrete: what must a rival still earn, obtain or accumulate before customers can choose an equivalent outcome?

MoatDefender: current methodology and limitations

Sources

Sources support the linked claims; our interpretation and checklists establish no outcome for your company.

Examine the evidence

Read the methodology Inspect the example Read the FAQ Does proof of precedence establish ownership?What is an original mind moat?