Archify gained 737 GitHub stars in one day, but the more revealing number is 9/9. The open-source agent skill has 11 checked-in examples in its public gallery, and each visible artifact carries a nine-check pass receipt. That pairing explains the interest better than another promise that an AI agent can draw a neat box-and-arrow diagram: Archify is trying to make the output inspectable before somebody treats it as a map of a real system.
The timing matters. Archify released version 3.0.1 on September 28, one day before the recorded star jump. It runs as a skill for coding agents including Codex CLI, Claude Code, Cursor, GitHub Copilot, and OpenCode. A user can describe a system in plain language or ask an agent to inspect a repository, then receive a standalone HTML file with an inline SVG diagram. The file can be opened locally and shared without installing Archify on the reader's machine.
The receipt is part of the artifact
The project's useful idea sits between the prompt and the picture. An agent first writes a typed JSON intermediate representation, or IR, for one of five diagram types: architecture, workflow, sequence, data flow, or lifecycle. Archify then checks that input against its schema and layout rules before producing the HTML. Its documented delivery flow says an invalid candidate cannot replace the last known good output.
That makes a failed render actionable. The validate --json and deliver --json commands return stable rule codes and identify the object that failed. They also list supported repairs, giving an agent a bounded correction target instead of a Node stack trace and a vague instruction to try again. The basic path is visible in the repository's command examples:
node archify/bin/archify.mjs validate workflow flow.json --json
node archify/bin/archify.mjs deliver workflow flow.json flow.html --json
Delivery is atomic, according to the Archify skill contract. The tool renders a candidate beside the intended output, checks it, and replaces the target only after the checks pass. In preview mode, an incomplete save leaves the previous verified diagram on screen. This is familiar behavior for anyone who has relied on a last-good build while editing a configuration file, applied here to an artifact that is usually judged by sight.
The Proof Lab makes those claims inspectable. It publishes the JSON input, finished HTML, graph size, short SHA-256 identifier, and the 9/9 result for examples such as an incident runbook and an agent lifecycle. These are maintainer-produced fixtures rather than an independent benchmark. They still give a developer something more useful than a screenshot because the input and output can be opened side by side.
Version 3 checks the source, too
A polished diagram can be wrong while passing every geometry test. Archify addresses part of that problem with optional repository evidence. When requested, a node can point to a file and line range at a specific Git commit. The validator checks the repository origin, commit, blob, path containment, and line bounds before accepting the reference. Version 3.0 extended that evidence model from architecture diagrams to workflows, sequences, data flows, and lifecycles.
This does not prove that the diagram captures the whole system. It proves a narrower and useful fact: the cited location exists in the named revision and meets the tool's reference rules. Archify's viewer labels evidence-backed nodes with source markers, while ordinary diagrams remain source-free. The repository's evidence description draws that line so a sketch made from a conversation does not look verified against code.
Version 3 also changes how dense diagrams are read. Lifecycle schema v2 places states on a shared grid and routes transitions through measured tracks. Sequence views can report unused horizontal space, and source badges now reserve room instead of covering labels. The 3.0 changelog says the viewer will fit a whole diagram onto the first screen down to a readable text floor, with zoom available when details need more room.
Those improvements have costs. The release embeds fixed JetBrains Mono font subsets so covered characters render consistently offline. The changelog puts the added weight at about 96 KB for each standalone HTML or SVG artifact. CJK glyphs may still fall back to fonts on the reader's platform, so identical bytes do not guarantee identical typography everywhere.
Why this is an agent skill
Archify delegates composition to the coding agent. The agent decides hierarchy, spacing, routes, and emphasis, then the local runtime checks the result. After generation, a user can ask for a focused change such as adding authentication or tracing a cache-miss path. The skill contract keeps the typed source available for the next edit rather than discarding it after one render.
Installation uses one command:
npx skills add tt-a1i/archify -g
The project's README also offers a one-shot skills use command for trying it without a permanent install. Our review of Archify covers the setup reality and where it fits beside other diagram tools. The repository uses the MIT license, including attribution to the earlier Cocoon AI project on which the skill says it is based.
Its boundaries are unusually explicit. Archify is not a general drawing editor, and hosted sharing or WYSIWYG editing is outside the current scope. The skill can accept pasted Mermaid through an agent, but the project does not present itself as an automatic Mermaid parser. The published scope instead keeps the output as typed source plus a self-contained viewer, favoring reviewable technical diagrams over free-form illustration.
There is one network detail worth noticing before a global install. Archify may fetch a fixed release manifest to display an update reminder. The project says that this check sends ordinary HTTP metadata but no prompt, project data, account identifier, or installed version. It does not install an update. Setting ARCHIFY_UPDATE_CHECK_DISABLED=1 disables the request and its reminder-state writes.
Where the checks stop
The word "verified" needs a tight boundary here. Archify's optional deployment-ownership profile can require named owners, regions, private database placement, and labeled boundary crossings. The repository documentation also says the profile does not inspect live infrastructure. A passing diagram confirms that required fields and relationships satisfy its authored rules; it does not confirm that a cloud account matches the picture.
Visual checking has a similar handoff. The project's visual-check command measures four desktop viewport sizes and captures light and dark evidence, but its roadmap leaves the final visual review pending. The viewer implements mobile controls and reduced-motion behavior. Its exports are deterministic, yet a machine receipt cannot decide whether a diagram teaches the right idea to its audience.
That limit is also the reason the project is interesting. Most diagram generators stop when they produce something plausible. Archify records the mechanical checks, preserves the old file when an answer fails, and states which judgments still belong to a person. The 737-star day records demand for that approach. It cannot prove that every generated map is correct.
Next, watch whether teams commit the typed JSON and receipts beside their code, then regenerate diagrams as repositories change. Commit-pinned evidence makes an old diagram an accurate historical snapshot, but freshness remains a workflow problem. The useful follow-up to a 737-star day is a duller measurement: how often real diagrams fail Archify's checks, and how often those failures stop a misleading artifact from being shared.