What Is ArtCraft? How AI Agents built 7 Adobe alternatives in 2 weeks

ArtCraft is a suite of 7 free, open-source applications built to replace Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat. What makes ArtCraft stand out isn't just that it's a free alternative, but how it was made: about 77% of its commits list Claude Opus 5.5 as a co-author, and the entire suite appeared in less than two weeks.

For the tech community, ArtCraft is a rare, public case study of AI agents writing software at scale. The project raises important questions: How complex can the software AI builds get? Does faster development mean professional-grade quality? And what do businesses need to prepare before bringing AI agents into their software development process?

What is ArtCraft? 7 open-source Adobe alternatives

ArtCraft is a project by Storyteller.ai (the storytold organization on GitHub), led by Brandon Thomas. In early October 2026, the team released ArtCraft Crafting Apps, 7 applications written from scratch in Rust:

The apps are free, run entirely on your machine, require no account or subscription, and support macOS, Windows, and Linux, with some also running in the browser. The team has since released WordCraft, GridCraft, and DeckCraft to replace Word, Excel, and PowerPoint, and says it is targeting AutoCAD and SolidWorks next.

Why is ArtCraft drawing so much attention from the tech community?

Interest in ArtCraft comes from two factors: demand for software that doesn't depend on subscriptions, and the speed of AI-assisted software development.

First, ArtCraft is a free alternative to Adobe

Adobe Creative Cloud offers a professional ecosystem of creative tools, but subscription costs can be a barrier for individuals, independent creators, and small businesses.

ArtCraft takes a different approach: free, open-source applications that run locally. This gives users more options, especially those who work on Linux or want to reduce their dependence on subscription services.

Still, cost is only part of the picture. Professional users also need file format compatibility, a plugin ecosystem, performance, and stability in their day-to-day workflows.

Second, and more important for the tech world: AI coding agents are shortening development time

ArtCraft has become a concrete example of a broader trend: AI coding agents no longer just help write individual snippets of code. They can take part in the software development process itself, from picking tasks and implementing features to testing and updating documentation.

A suite of applications that took Adobe decades to build had a "working" version in under two weeks. The story quickly became the clearest example in a heated debate: when AI writes code well enough, how high is the barrier to building competing software?

Brandon Thomas himself declared on Hacker News: "Everything can be open source multiplatform Rust now."

How was ArtCraft built with Claude and AI agents?

What makes ArtCraft unusual is that its development process can be observed through its source code, commit history, and the AI agent guidelines published on GitHub.

Instead of handing the AI a broad request like "build a photo editing app", the team set rules, broke the work into small tasks, defined clear completion standards, and built a rigorous testing process.

Key numbers on AI coding in ArtCraft

• 5,031 commits across 7 repos, of which 3,876 (about 77%) list Claude Opus 5.5 as co-author
• The rest involve Claude Sonnet 5.5, Claude Fable 5.1, OpenAI Codex, and GitHub Copilot
• In PhotoCraft alone, commits co-authored by Opus 5.5 added more than 244,000 lines of code
• Public history starts on September 30–October 1, 2026, and the apps launched in the first week of October

The team's process can be broken down into the five steps below.

Step 1: Set rules for AI agents before writing any code

Each repo has its own guidelines file for AI agents, such as AGENTS.md (some repos also include CLAUDE.md). Agents must read these files before doing anything else. Some notable rules:

• Clean-room: agents may only learn the behavior and look of professional software through public documentation, published shortcut lists, or by using the app. They must not look inside Adobe's software ("Never inspect application internals") or copy Adobe's code, shaders, or assets.
• Never crash the app: code must not panic and must return errors instead (proper error handling), and every crash fix must come with a regression test so the bug doesn't come back.
• Every feature must be a "command" with its own ID and documentation. This way, the UI, the command line, and AI agents all invoke the same feature in the same way.
• Every image and asset must have a clear license and be attributed in the same change.
• Technical constraints, such as no Electron or web frameworks for the UI.

Step 2: Break goals down into testable tasks

Agents don't come up with work on their own. The team maintains:

• A roadmap and list of milestones
• A prioritized list of gaps compared with Adobe's software
• A scorecard of which features are complete, partial, or missing
• An automated check that lists which menu items from the original software are still missing

Priorities are spelled out: infrastructure first, then easy wins, then everything else.

Step 3: Agents work in a fixed loop

For each task, an agent goes through five steps:

1. Orient: pick the next open task and read the architecture docs for the relevant code
2. Write tests first: "Write down the acceptance test before writing code", defining what "done" means before writing any code
3. Implement and run tests
4. Verify: run CI (automated checks covering formatting, linting, and tests). For UI features, the agent launches the in-development app in control mode, clicks and drags through the feature itself, and takes screenshots to check the result
5. Record: update docs, the roadmap, and the dev log so the next agent knows where things stand, then commit with the task ID and move on

One notable detail: to test video decoding, FilmCraft uses FFmpeg (a widely used video tool) as a test oracle, comparing the app's output against FFmpeg's, rather than using FFmpeg inside the app.

Step 4: A clear definition of done

A feature only counts as done when it meets a full checklist: registered as a documented command, covered by tests (including undo/redo), wired to menus and shortcuts, driven by an agent with screenshots reviewed, passing CI, and with docs updated.

This standard is what keeps thousands of AI-written commits at a minimum level of quality without anyone reading every line of code.

Step 5: Agents run in parallel, humans keep the key decisions

Multiple agents can work at the same time. Each agent works in its own git worktree, and each crate (Rust module) has exactly one agent owner, so agents don't overwrite each other's code.

Agents are told "Don't block on choices you can make yourself", meaning they decide whatever they can on their own. They only stop to ask a human team member when it comes to:

• Licenses outside the allowlist
• Spending money
• Publishing or pushing code to the main repo
• Personal data
• Destructive git operations, such as force pushes
• Anything the rules don't clearly cover, following the principle "when in doubt, leave it out"

Tasks like overall planning, releases, and signing installers remain in human hands.

In short: humans are no longer the ones writing code. They write the rules, divide the work, set the standards, and approve risky decisions. AI agents handle the rest.

Why are the apps designed to be controlled by AI?

Because every feature is a "command", ArtCraft apps can be controlled through the command line (CLI), a JSON control channel, or an MCP server. This design serves two purposes:

• During development: it's how AI agents open the app, operate it, and verify the features they just built (Step 3 above)
• For end users: MCP-enabled AI tools such as Claude Desktop or Cursor can also control the apps, for example batch color-correcting photos according to a rule

In other words, the software is built so that AI can write it, test it, and use it.

Can ArtCraft replace Adobe yet?

The short answer: not yet. ArtCraft shows that AI coding agents can help build applications with substantial feature coverage in a short time, but it hasn't proven it can fully replace Adobe Creative Cloud for professional work.

The team is candid in its own assessment. PhotoCraft's README states the app is "not yet a Photoshop replacement for daily professional work", estimating it meets only about 25–35% of professional daily needs. FilmCraft rates itself at about 86% feature breadth but only about 56% "ready for real work" (measured against Premiere Pro on October 10, 2026). Early testers reported lag, broken text editing, and unresponsive shortcuts.

AI shortens the time to write code, but not the time to prove reliability

This is the most important lesson from ArtCraft. Three very different things tend to get lumped together:

1. Explaining how a piece of software works
2. Rebuilding a feature
3. Reliably replacing a professional workflow

AI is doing the second faster than ever. But the third still takes time, real data, and testing under real conditions. "627 of 627 menu items wired to commands" only shows that commands are connected to the app, not that features behave correctly. Professional users need to know whether layered files survive repeated edits, whether exports meet print requirements, and whether projects reopen months later.

Put simply: the cost of attempting to replace software is falling fast, but the cost of proving it's reliable is still high.

Clean-room development: The legal boundaries of AI-rebuilt software

ArtCraft describes itself as a "clean-room" project: contributors may only observe the behavior and look of Adobe's software, and code must be written from public technical documentation, never from Adobe's source code.

This stands out next to REA, another viral tool that lets AI agents run reverse engineering tools such as Ghidra or IDA on closed-source software and rebuild its functionality. Brandon Thomas reportedly refuses to use REA and doesn't allow decompiling commercial software in the project, out of concern that AI could inadvertently reproduce proprietary code and invite litigation.

Still, open questions remain:

• Adobe's terms of use prohibit reverse engineering, although laws in some countries allow software analysis for interoperability
• It's unclear whether any features overlap with Adobe's patents
• The clean-room process currently rests on the project's internal rules and hasn't been independently verified

At the time of writing, Adobe has not issued any official response.

For businesses, this is a practical reminder: when AI agents write your code, the question "where did this code come from?" matters as much as "does it work?". Processes and rules for agents should be set from the start, not after problems arise.

4 lessons businesses can learn from ArtCraft

1. The cost of building software is falling, and competitive advantage is shifting. If a small team can build a functional copy of major software in a few weeks, lasting advantage lies in what's harder to copy: data, ecosystem, reliability, support, and customer trust.

2. Agents are only effective with guardrails. What let ArtCraft move fast wasn't just a powerful model, but pre-written rules for agents, a clear task list, tests written first, a concrete definition of done, and mandatory checkpoints for human approval. Without these guardrails, autonomous agents can cause unpredictable damage.

3. "Agent-native" software will become increasingly common. ArtCraft designs every function so AI agents can call it via CLI, a JSON control channel, and MCP. It's a sign that future software will need an interface for AI, not just for humans.

4. No single model does everything well. Even in ArtCraft, Claude Opus 5.5 handled most of the work, but the project still combined Sonnet, Fable, Codex, and Copilot. Choosing the right model for the right job, balancing quality, speed, and cost, is becoming a core skill for engineering teams.

Frequently asked questions about ArtCraft

Is ArtCraft free?

Yes. The Crafting Apps are free and open source. ArtCraft makes money from a separate product, ArtCraft IDE, an AI image and video creation tool that's free to download, with credits charged when calling AI models.

Which AI was used to build ArtCraft?

According to public commit data, about 77% of commits list Claude Opus 5.5 as co-author. The project also used Claude Sonnet 5.5, Claude Fable 5.1, OpenAI Codex, and GitHub Copilot.

Can ArtCraft replace Adobe?

Not yet. The apps are in alpha. The team itself says PhotoCraft can't yet replace Photoshop for daily professional work.

What does "clean-room" mean?

It's a method of rebuilding software from scratch based on observed behavior and public documentation, without copying or decompiling the original source code. The goal is to reduce the risk of copyright infringement.

Should businesses use AI agents to build software?

AI agents can significantly shorten development time, but they need rigorous testing, clear rules on code provenance, and human approval points for important decisions.

Choose the right model for the right job

ArtCraft shows that AI can now create not just content, but entire software products. But results don't come from a single "most powerful" model. They come from combining multiple models, each handling the work it's best suited for.

If you want to test and compare multiple models in one place,  GreenNode Model-as-a-Service  offers models from multiple providers, including Google, Anthropic, OpenAI, and GreenNode, through a single API. You can try them directly in the Playground to compare quality, latency, and cost before integrating them into your system.

Explore models on GreenNode MaaS →