Let's Open Source the World with Brandon Thomas, ArtCraft

AI-driven development now makes it possible to fully recreate and open-source entire categories of proprietary, extractive software (Adobe Creative Suite, Mi

By Sean Weldon

Let's Open Source the World: Agentic Replication of Proprietary Creative Software

Abstract

This synthesis examines a claim advanced by Brandon Thomas of Artcraft that agentic artificial intelligence (AI) development has lowered the cost of software replication sufficiently to make wholesale open-source reconstruction of proprietary application categories tractable. The case study is Artcraft's reported recreation of Adobe Photoshop, InDesign, Premiere, Animate, Acrobat, Figma, and core office productivity applications through structured prompting and agent-driven implementation in Rust. Evidence presented includes a functional Photoshop approximation produced within 24 hours, feature parity advancing from 50% to above 60%, roughly 1.7 million daily downloads, 300,000+ GitHub stars, and 4,000 contributors. The analysis situates these claims within a historical argument about the enclosure of the open web, extracts the architectural and operational decisions that made the approach viable - MCP/CLI-first design, self-hosted continuous integration (CI), and a differentiated human-oversight policy - and identifies unresolved questions around long-tail parity and agent supervision at scale.

1. Introduction

The thesis under examination holds that AI-driven development now permits the full recreation and open-sourcing of entire categories of proprietary software, reversing the consolidation and gatekeeping characteristic of the Web 2.0 era. This claim is distinct from the familiar observation that large language models (LLMs) accelerate individual coding tasks; it asserts that the economic moat protecting mature commercial applications - feature depth accumulated over decades - is now surmountable by a comparatively small team coordinating agent fleets against production-quality targets.

Two terms anchor the analysis. Agentic development denotes software construction in which an AI agent plans, implements, and iterates against a specification with bounded human review, rather than producing isolated completions. Feature parity, as used in the source material, means functional equivalence with a reference application's documented toolset rather than visual or codebase similarity, and is explicitly tracked as a percentage metric that increases over time.

The analysis proceeds as follows. Section 2 situates the project within the historical narrative of platform enclosure that motivates it. Section 3 examines the replication methodology, community growth, and architectural choices. Section 4 isolates transferable technical findings. Sections 5 and 6 assess implications and open questions, including the credibility gap between self-reported metrics and independently verifiable outcomes.

2. Background and Related Work

The project's stated motivation rests on a historical contrast between the early web's protocol-centric openness - open APIs on Reddit, Flickr, and Twitter; user-editable HTML/CSS on MySpace; owned rather than licensed media - and the subsequent consolidation associated with Web 2.0, in which "platforms replaced protocols" and mobile distribution closed off the open app ecosystem that existed on desktop. This trajectory is described using the term enshittification, the pattern in which a platform, having won market share, progressively degrades terms to extract rent, exemplified by Adobe's disclosed-in-fine-print $200 cancellation fee for subscriptions terminated early.

This framing is not incidental; it supplies the normative justification for treating software recreation as restorative rather than merely competitive. The relevant intellectual context is the long-standing tension between open protocols and closed platforms in internet history, and the project positions AI-assisted development as a novel lever capable of tipping that balance back toward openness after roughly two decades of platform ascendancy.

3. Core Analysis

3.1 Replication Methodology

The reported starting point was a single structured prompt instructing an AI system on the "basic shape" of Photoshop, explicitly separating UI/UX design, enforcing MCP and CLI compatibility, defining per-tool behavior plans, data structures, and file-format importers. Thomas reports that within 24 hours this produced a functional approximation of Photoshop that felt familiar to users and ran approximately 10x faster than the original. Feature parity is described as having progressed from an initial 50% to "past 60%" at time of reporting, with an expectation of approaching near-complete parity following a Pareto principle trajectory - rapid gains on the most-used 20% of functionality followed by a long tail of diminishing-return features. The team reports deliberately prioritizing fidelity to Adobe's existing interface and interaction model over designing novel agentic interfaces, citing preservation of artists' muscle memory and a desire to first demonstrate parity before innovating further. The same methodology was reportedly extended to InDesign, Premiere, CAD tools, Flash, digital audio workstations, Microsoft Office equivalents, Google Docs equivalents, and Figma - the latter including a server-side multi-user collaboration backend built on Raft/Paxos consensus algorithms. Additional reported outcomes include localization into languages unsupported by original vendors (e.g., Arabic in InDesign) and continued support for vendor-discontinued products (Adobe Animate) at user request.

3.2 Community Scale and Adoption

The reported growth metrics are substantial: a Discord community that grew from 3,000 members to a size exceeding the combined Discord membership of three named AI tool communities (Jenni, Suno, Higgsfield); coverage by Reddit and the New York Times within two days of initial virality; more than 300,000 GitHub stars; 4,000 contributors; roughly 1.7 million downloads per day with accelerating growth; and a reported 300 community-built applications. A quarter of users are described as running Linux, a platform historically underserved by professional creative and office software. Thomas also reports a qualitative shift among previously anti-AI artists, who are described as becoming users and contributing feature requests and bug reports.

3.3 Human-in-the-Loop Policy

A central operational finding is a differentiated oversight policy: human review is applied to backend work involving schema migrations and API design, while agents are permitted to work with minimal supervision on "immutable" surfaces such as desktop applications and games. This is illustrated by a reported case of a community member running 40 machines to rebuild Unreal Engine in Rust largely unsupervised by human review. This policy functions as a risk-stratification heuristic: irreversible or shared-state changes receive gating, while self-contained, easily discarded or regenerated work does not.

4. Technical Insights

Several implementation decisions are presented as load-bearing rather than incidental.

A trade-off worth flagging: the "99.99% parity" target is explicitly framed as a long-tail problem under the Pareto principle, meaning the reported 50%→60%+ progress likely reflects the comparatively easy initial gains, and the claim's generalizability to the full long tail of professional workflows remains unverified in the source material.

5. Discussion

The central implication, if the reported metrics hold under independent scrutiny, is that agentic development may compress the capital and headcount historically required to reach competitive feature parity with mature commercial software, altering the economics that have sustained platform consolidation since the Web 2.0 transition. This would extend arguments about AI-driven productivity gains from individual developer tasks to category-level market structure. The explicit application of a human-in-the-loop policy stratified by reversibility, rather than by task type alone, is a transferable governance pattern independent of whether the broader open-source thesis proves durable.

However, the source material is a single self-reported case study, presented by a project's own founder, with metrics (downloads, stars, parity percentage) that are not independently audited within the material itself. Feature parity in particular is a notoriously difficult metric to define rigorously for applications like Photoshop or Premiere, which carry decades of edge-case behavior, plugin ecosystems, and color-management or file-format compatibility requirements that are not easily captured by a percentage figure. The claim of "99.99%" eventual parity is aspirational rather than demonstrated. Future investigation would benefit from independent benchmarking against documented toolsets, scrutiny of the long-tail features, and longitudinal tracking of whether community contribution volume is sustained once initial virality subsides.

6. Conclusion

The case study documents a specific methodology - specification-first agentic prompting, Rust as implementation substrate, differentiated human oversight by reversibility, and self-hosted CI to absorb contributor-driven scale - applied to the recreation of multiple categories of proprietary creative and office software. The reported outcomes (rapid initial functional parity, large contributor and download figures, and extension across product categories) are significant if verified, and the operational lessons around CI scaling and oversight policy are directly applicable to other large-scale agentic open-source efforts regardless of one's assessment of the broader "open source the world" thesis. The unresolved question is whether long-tail feature parity and contributor engagement are sustainable at the claimed pace once novelty-driven growth subsides.


Sources


About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub

Sean Weldon builds apps and custom software at 22Atlas.

Start your build

More about Sean