AI Software Clones in 2026: Can AI Coding Agents Rebuild Photoshop and Microsoft Office?

What the ArtCraft and REA projects reveal about the new economics of software development—and why a working demo is not the same as a professional replacement.

In late September and October 2026, a wave of AI-assisted software projects captured developers' attention. One developer began publishing open-source applications inspired by major creative tools, then extended the effort to Word-, Excel-, and PowerPoint-style software. Around the same time, Reverse Engineer Anything (REA) gained attention for connecting coding agents to established reverse-engineering tools.

The headline is tempting: if one developer can use AI agents to recreate parts of Photoshop or Microsoft Office, is traditional desktop software about to become obsolete?

The evidence points to a more interesting conclusion.

AI coding agents are lowering the cost of building broad software prototypes and compatible alternatives. They have not eliminated the hard parts of software engineering: fidelity, reliability, file compatibility, performance, accessibility, security, and years of accumulated workflow details.

The distinction matters for users choosing software, developers looking for product opportunities, and companies deciding where to invest.

The short answer

• Can AI agents help recreate complex applications? Yes. Projects such as ArtCraft show that an individual developer can coordinate agents to build substantial, multi-component applications faster than a traditional solo workflow might allow.
• Can these projects replace Photoshop or Microsoft Office today? Not across the board. Their own project documentation still identifies important gaps, and readiness estimates differ by application and user type.
• What has changed? The cost of implementing a wide feature surface is falling. The cost of proving that those features behave correctly remains significant.
• Where is the opportunity? In focused alternatives, interoperability tools, automation, specialized workflows, and software that is designed to be controlled by both people and agents.

1. What happened: ArtCraft is rebuilding familiar software workflows

ArtCraft is a family of open-source applications written in Rust. Its projects aim to recreate familiar creative and productivity workflows using independently implemented code.

Examples include:

• PhotoCraft: a Photoshop-style image editor.
• WordCraft: a Word-style word processor.
• GridCraft: an Excel-style spreadsheet.
• DeckCraft: a PowerPoint-style presentation tool.

The projects describe themselves as clean-room reimplementations based on public specifications and observed behavior. They are not official Adobe or Microsoft products, and their relationship to those products should not be confused with endorsement or affiliation.

The technical ambition is notable. These are not merely landing pages with a few AI-generated buttons. The repositories describe document models, rendering systems, command registries, file-format support, tests, desktop applications, and browser builds.

The project details also show why the story needs nuance. The published readiness assessments distinguish between having a feature represented in the interface and having that feature work reliably enough for professional use.

For current details, start with the official repositories: PhotoCraft, WordCraft, GridCraft, and DeckCraft.

2. REA is not the same thing as ArtCraft

The two projects are often discussed together, but they solve different problems.

ArtCraft is a set of applications. REA is a toolkit that helps coding agents investigate software.

REA connects compatible AI agents to tools for inspecting applications, JavaScript and Electron packages, .NET assemblies, and native binaries. Depending on the target and installed providers, it can help collect structural and behavioral evidence that an agent can use to understand a feature or guide an independent implementation.

REA does not automatically turn any commercial program into a perfect clone. Its official documentation emphasizes evidence, limitations, and authorized use. Native analysis may require separate tools such as Ghidra, Hopper, or IDA, each with its own licensing terms.

Official references:

• REA on GitHub
• REA responsible-use disclaimer

This distinction is important. A reverse-engineering tool can make investigation more accessible; an application reimplementation still requires architecture, implementation, testing, integration, and ongoing maintenance.

3. Why can a small team build so much software with AI?

The breakthrough is not that an AI model can write every line perfectly. It is that several expensive parts of development can now be accelerated and coordinated.

A. Parallel work on independent components

A complex editor can be decomposed into components such as:

• Document and file-format handling.
• Rendering and compositing.
• Editing commands.
• Menus, panels, dialogs, and keyboard shortcuts.
• Import/export.
• Automated tests.

Coding agents can investigate or implement separate components in parallel when their interfaces are clear. A human engineer still needs to define the architecture, resolve incompatible assumptions, review the output, and integrate the result.

Parallelism helps most when tasks are independent. It creates less benefit—and may add coordination overhead—when every task depends on the same central data model.

B. Existing formats and public specifications provide a starting point

Many mature applications use documented or partially documented file formats. Developers can build independent readers and writers, then test them against sample files.

For example, PhotoCraft documents support for PSD and other image formats, while WordCraft and GridCraft focus on document and spreadsheet formats familiar to users of existing office software.

File compatibility is a particularly valuable target because users often care more about whether their files open, preserve formatting, and round-trip correctly than about the exact internal implementation.

C. A shared command layer makes software easier to automate

One of the most interesting architectural choices in these projects is that user-interface actions can also be represented as commands.

A command can potentially be invoked by a button, keyboard shortcut, command-line interface, automation script, or an AI agent through an interface such as MCP.

This creates a more general form of software: the application is not only interactive for a human, but also programmatically operable.

D. Testing turns generated code into engineering progress

AI-generated code is cheap to produce, but a plausible implementation is not proof of correctness.

The most valuable workflow is a loop:

1. Define a feature and expected behavior.
2. Implement it.
3. Run automated tests.
4. Compare output against known results.
5. Fix the mismatch.
6. Repeat.

For an image editor, this can include rendering comparisons and file round-trip tests. For a word processor, it can include document layout, formatting preservation, and real-world DOCX files. For a spreadsheet, it can include formula correctness, calculation dependencies, and XLSX round-tripping.

The key point is that AI agents can accelerate implementation, but the quality of the test oracle—the method used to determine whether the output is correct—often determines how much progress is real.

4. How close are the projects to Photoshop and Office?

The official project scorecards provide a useful reality check. The figures below are project self-assessments, not independent laboratory benchmarks. They also use different target applications and scorecards, so they should not be interpreted as a perfectly standardized league table.

Their assessments still identify gaps in file fidelity, layout, stability, performance, professional workflows, and plug-in compatibility. These figures come from different project scorecards and are not independent benchmark results.

Snapshot dates: PhotoCraft assessment updated October 18, 2026; WordCraft assessment updated October 16; DeckCraft assessment updated October 14. The projects change quickly, so consult their current scorecards before relying on the numbers.

Sources: PhotoCraft parity assessment, WordCraft parity assessment, and DeckCraft parity assessment.

Why feature coverage is not the same as parity

Consider an image editor with a menu item for every tool in Photoshop. That is useful progress, but it does not mean every tool behaves the same way, handles every edge case, or performs well on a large production document.

The same applies to office software:

• A spreadsheet may implement hundreds of functions but still mishandle a complicated workbook.
• A word processor may support many formatting controls but paginate a long document differently.
• A presentation editor may expose animations but alter timing or rendering when opening a real deck.
• An image editor may open a PSD file but lose effects, text behavior, or compatibility when saving it back.

Professional software is defined by the consistency of the entire workflow—not just the number of visible features.

What about GridCraft?

GridCraft describes support for more than 500 worksheet functions, XLSX-oriented workflows, charts, pivot tables, and agent-driven commands. That is a substantial feature surface.

But a function count does not by itself establish Excel compatibility. Users need to test representative workbooks containing formulas, styles, charts, external links, named ranges, pivot tables, and edge cases relevant to their work.

For a developer evaluating GridCraft, the right question is not "Does it have 500 functions?" It is "Does it preserve and calculate the workbooks my users actually depend on?"

5. Why familiar workflows matter more than a long feature list

There is a reason people hesitate to switch from Photoshop, Excel, or Word even when free alternatives exist: switching has a cost.

Users have learned keyboard shortcuts, muscle memory, file conventions, templates, plug-ins, and collaboration practices. An alternative that reproduces familiar interactions can reduce that learning cost.

However, matching a visual layout is only the beginning. Users also expect:

• Predictable keyboard shortcuts and selection behavior.
• Stable undo and recovery.
• Consistent typography, pagination, and rendering.
• Reliable import and export.
• Accessibility and localization.
• A mature extension ecosystem.
• Long-term support and compatibility with colleagues' files.

These expectations accumulate over years. AI can help implement them, but it does not remove the need to verify each behavior.

6. Does "clean-room" mean the software is automatically legal?

No.

A clean-room approach generally attempts to separate observations or specifications about behavior from the independent implementation. It can be relevant to compatibility work, but the label itself is not a legal ruling.

The legal position depends on the jurisdiction, the particular software and assets involved, the method used to study it, the license and contractual terms, and the way the resulting product is distributed or marketed.

Issues may include:

• Copyright in source code, artwork, icons, text, and other protected expression.
• Trademarks and branding that could imply affiliation.
• Interface elements that may receive different forms of protection depending on jurisdiction.
• Patents and other applicable rights.
• Contractual restrictions or anti-circumvention rules.
• License obligations for third-party libraries, fonts, and assets.

The safest general principles are to use public specifications where appropriate, create original assets and implementation, respect third-party licenses, avoid bypassing DRM or license checks, and obtain authorization before analyzing software when required.

REA's own responsible-use disclaimer explicitly excludes bypassing license checks and copying proprietary code or assets. Developers should read the relevant project licenses and seek qualified legal advice before commercializing a product that closely targets another application.

This section is general information, not legal advice.

7. What this means for the software industry

The immediate implication is not that established software vendors disappear overnight. It is that the economics of launching a credible alternative are changing.

More prototypes will reach the market

A solo developer can use coding agents to create a broader first version than would previously have been practical. That increases experimentation and may bring niche tools to users who are poorly served by large suites.

It also means a larger number of competing projects. A generated application is not automatically a sustainable business: distribution, reliability, support, and user trust still matter.

The advantage shifts toward focused products

Rebuilding every feature of a large creative suite is expensive even with AI. A smaller product that solves one high-value problem can be easier to test, ship, and maintain.

Examples include:

• A batch image-processing tool designed for one professional workflow.
• A spreadsheet cleanup and validation utility.
• A presentation accessibility checker.
• A document conversion tool that preserves a specific set of formatting features.
• A PSD or XLSX inspection utility for developers and QA teams.

These are examples of product directions, not claims that the markets are underserved. Each needs validation against existing alternatives and real search demand.

Agent-native software may become a competitive advantage

An application with stable, documented commands can be automated by scripts and AI agents. This can make it useful in workflows that do not require a person to click every button.

For example, an image editor might expose a controlled operation to resize a folder of images, apply a filter, export files, and return a machine-readable report.

That capability can be more valuable than merely adding a chatbot to the interface.

8. Product opportunities for indie developers

If you are building small software products with AI coding agents, the best lesson is not to attempt a complete Photoshop clone. It is to identify a narrow workflow that users need repeatedly.

Before choosing one, use a repeatable opportunity test:

1. Find a clearly defined user problem.
2. Search for existing products and open-source alternatives.
3. Inspect user complaints and feature requests.
4. Check Google search demand and competition.
5. Build a small prototype around one core workflow.
6. Test with real files and users.
7. Expand only after the core task is reliable.

This is where AI-assisted development becomes especially powerful: it can reduce the cost of testing a product hypothesis. It cannot tell you, by itself, whether enough users want the product.

9. How to evaluate an AI-built software alternative

If you want to try PhotoCraft, WordCraft, GridCraft, or DeckCraft, do not rely only on screenshots or a short demo.

Use a small, representative test suite.

For an image editor

• Open layered files with masks, text, and effects.
• Edit and save them, then reopen the results in both applications where appropriate.
• Compare rendered output.
• Test undo, large images, and common export formats.
• Check missing AI features and plug-in compatibility.

For a word processor

• Open real DOCX documents with tables, images, headers, footnotes, and tracked changes.
• Compare page breaks and layout.
• Save and reopen the document in the original application.
• Check fonts, special characters, and print/PDF output.

For a spreadsheet

• Use workbooks with formulas, charts, named ranges, formatting, and multiple sheets.
• Compare calculated values with a trusted reference.
• Save and reopen in Excel or another target application.
• Check whether unsupported features are preserved or silently lost.

For presentation software

• Test existing decks with complex layouts, fonts, charts, animations, and embedded media.
• Compare slides after round-tripping through the alternative.
• Test presenting, exporting, and switching between displays.

For every test, record what worked, what changed, what failed, and whether the failure was visible to the user. A compatibility report is much more useful than a single overall percentage.

Final verdict: software is not over, but software development is changing

ArtCraft and REA illustrate two sides of the same shift. Coding agents can help developers understand complex systems and implement broad feature sets more quickly. The resulting projects can be real, substantial software—not just mockups.

But the gap between a compelling prototype and a dependable professional application remains large. Compatibility, correctness, stability, performance, accessibility, security, and long-term maintenance are still engineering problems.

For users, these projects are worth watching and testing on non-critical work. For developers, they are evidence that the cost of experimentation is falling. For businesses, they are a reason to examine which parts of their software are genuinely differentiated and which parts are primarily expensive to maintain.

The most promising opportunity is not necessarily a one-person replacement for every major software suite. It may be a collection of smaller, more focused applications that solve specific problems, work with existing files, and expose their capabilities to both humans and AI agents.

The next generation of software may not be defined by how many features an application has. It may be defined by how reliably those features work—and how easily they can be used, tested, and automated.

Frequently asked questions

Can AI really rebuild Photoshop?

AI coding agents can help implement substantial parts of an image editor, including layers, tools, filters, and file handling. PhotoCraft's own assessment still reports a meaningful gap between feature breadth and readiness for professional work, especially around file fidelity, performance, AI features, and ecosystem compatibility.

Can AI-built alternatives replace Microsoft Office?

Some workflows may already be useful for experimentation or limited tasks, but real-world readiness varies by application. Users should test their own documents, spreadsheets, and presentations before relying on an alternative for critical work.

What is REA?

REA is an open-source toolkit that connects AI coding agents to reverse-engineering and application-analysis tools. It helps agents collect evidence about software behavior; it is not an automatic perfect-clone generator.

Are ArtCraft's applications free and open source?

The individual repositories publish their own licenses, including MIT and/or Apache-2.0 licensing for the projects discussed here. Check each repository, its dependencies, fonts, assets, and release terms before redistribution or commercial use.

Is it legal to clone an existing application?

There is no universal yes-or-no answer. Independent implementations and interoperability work may be treated differently depending on jurisdiction and circumstances. Copyright, trademarks, patents, contracts, and anti-circumvention rules may all be relevant. Obtain legal advice for a specific commercial project.

What should indie developers build with AI coding agents?

Start with a narrow workflow that is painful, repeated, and easy to validate. File conversion, document checking, batch processing, spreadsheet auditing, and specialized automation are possible directions—but demand and competition should be researched before investing.

---

Sources and further reading

Official projects

• PhotoCraft — open-source image editor
• PhotoCraft parity assessment
• WordCraft — word processor
• WordCraft parity assessment
• GridCraft — spreadsheet application
• DeckCraft — presentation application
• DeckCraft parity assessment
• REA — Reverse Engineer Anything
• REA responsible-use disclaimer

Reporting and analysis

• KDnuggets: The Rise of AI Software Clones
• Flowtivity: Someone Rebuilt Adobe in Rust with AI Agents. Is It Usable?
• Decrypt: Developer Who Cloned Photoshop With AI Just Did the Same to Microsoft Office

Editorial note: This article reflects publicly available project documentation and reporting checked on October 22, 2026. Project status, parity estimates, licenses, and capabilities can change quickly. Read the current repositories and license notices before downloading, distributing, or using these applications in production.