Archify Hit 58,000 Stars in Five Months and One Person Wrote 72% of the Commits
A JavaScript agent skill that turns code descriptions into interactive architecture diagrams gained 58,585 stars since April. We read the source, checked the contributor graph, and found a solo project that 3,820 people forked but almost nobody contributed back to.
The WJS Desk
Sep 12, 2026 · updated 12 days ago · 5 min read
Archify showed up on GitHub Trending in late August at around 4,000 stars. Two weeks later it was the number one trending repository. As of September 12, it sits at 58,585 stars with 3,820 forks, gaining roughly 3,900 stars per day at its peak. We pulled the contributor data, read the source, and checked whether the velocity matches the substance.
The short version: the diagram engine is genuinely well-built and the validation layer is something most diagram tools skip entirely. The star count is earned. The contributor graph is the part that should give you pause.
The number that made us look
58,585 stars in roughly five months since the repository was created on April 15, 2026. That is faster than VoiceStudio (22,000 in five months, which we covered last week) and puts it in the same trajectory tier as DeepSeek Harness earlier this year. The acceleration was not gradual. Archify sat under 5,000 stars until late August, then something caught and it added over 50,000 in under three weeks.
3,820 forks is unusually high relative to the star count (a 6.5% fork ratio versus the typical 2 to 4% for trending repos), which suggests people are actually downloading and modifying it rather than just starring for later.
What it actually is
Archify is an agent skill, not a standalone application. You install it into your AI coding agent (Claude Code, Codex CLI, Cursor, Raven, DeepSeek Harness, or OpenCode) and then ask the agent to diagram your system. The agent sends a description, Archify compiles it through a typed JSON intermediate representation, and out comes a self-contained HTML file with interactive SVG.
Five diagram types are supported:
- Architecture: Components, services, storage, and trust boundaries
- Workflow: CI/CD processes, approvals, tool calls, runbooks
- Sequence: API interactions, cache behaviour, async traces
- Data flow: Pipelines, lineage tracking, sensitivity boundaries
- Lifecycle: States, retries, waits, terminal outcomes
The output is a single HTML file with embedded SVG, dark/light theme toggle, animation, and export to PNG, SVG, or WebM. No external dependencies at runtime. You can drop the file into a README, a wiki, or a Slack message and it just works.
npx skills add tt-a1i/archify -gThe validation layer is the real product
Most diagram generators stop at "does the SVG render." Archify runs an atomic validation pipeline before delivering output: schema check, layout rules, HTML/SVG render verification, route validation, and label clearance (ensuring text does not overlap). If any check fails, it returns machine-readable diagnostics with specific repair guidance instead of a broken diagram.
This matters because the primary consumer is an AI agent, not a human. When a model generates a diagram description with an invalid edge or an overlapping label, Archify catches it and tells the model what to fix. Without this loop, you get the same garbage-in-garbage-out problem that plagues every other AI-to-visual pipeline.
The deterministic rendering is the other detail worth noting. The same JSON input always produces the same visual output. This sounds obvious, but most layout engines introduce nondeterminism through force-directed algorithms or random initialisation. Archify uses agent-chosen hierarchy instead, which means your diagrams do not rearrange themselves between runs.
What it does not do
- No automatic code analysis. Archify does not read your codebase and generate a diagram. Your AI agent reads the code and writes the description. Archify renders what the agent tells it. If the agent hallucinates a service that does not exist, the diagram will contain it.
- No real-time sync. Diagrams are snapshots. Change your code and the diagram is stale until you regenerate it.
- No collaborative editing. One agent, one output. There is no multiplayer mode or version diffing beyond the
compareCLI command. - No UML. The five diagram types are opinionated and none of them map to standard UML. If your organisation requires UML compliance, this is not the tool.
- No self-hosting of the preview. The
previewcommand runs on loopback only. There is no built-in way to share live previews with a team.
The accuracy ceiling is your agent, not Archify. The source evidence feature can pin diagram nodes to specific Git commits, but this only works when the agent correctly identifies the relevant files. We could not test this ourselves because it requires running inside an agent session with a target codebase, and the results would depend entirely on which model was driving. Treat the "SRC" pins as aspirational until you verify them in your own environment.
Who made it, and does that matter
One developer, tt-a1i, wrote 173 of 240 commits (72%). The next contributor, sunsunsun-java, has 31 commits. After that the numbers drop sharply: 7, 4, 3, 3, 2, 2, 2, and then fifteen single-commit contributors. Twenty-two people have contributed in total.
| Contributor | Commits | Share |
|---|---|---|
| tt-a1i | 173 | 72% |
| sunsunsun-java | 31 | 13% |
| All others (20) | 36 | 15% |
The codebase is JavaScript, MIT licensed, at 139 MB on disk. The current development version is v2.17.0-dev.1, which means the project has shipped a significant number of releases in five months. That velocity from one person is impressive and also the risk. If tt-a1i steps away, there is no obvious successor, and the 31-commit second contributor has not touched the validation layer or the core rendering engine based on the commit messages.
The MIT licence and 139 MB repo size mean vendoring it is technically possible. If you depend on Archify in your workflow, downloading a known-good version and pinning it locally is cheap insurance.
Verdict
Archify solves a real problem (architecture documentation decays the moment you draw it) with a genuinely clever approach (make the AI agent generate it, validate it programmatically, render it deterministically). The validation pipeline is the feature that separates it from the dozen Mermaid alternatives that appeared this year, and the self-contained HTML output is the distribution insight that made it viral.
If you use any AI coding agent and you need architecture diagrams, install it today. The tool works, the output is good, and the worst case is that you have a folder of standalone HTML files that render without any external dependency.
Adopt it, pin a version, and do not build a workflow that breaks if one person in Shenzhen loses interest.
If you are evaluating it for a team standard, wait until the contributor base broadens or until someone forks it with a second core maintainer. 58,000 stars is validation of the idea. It is not validation of the project's longevity.
