Archify draws architecture your agent cannot make up

A diagram skill built on a typed JSON IR with deterministic checks, so nodes trace to real source and two snapshots diff into added, removed and rerouted.

tt-a1i/archify
Stars
13K
Language
HTML
License
MIT
Age
7 months old
Last push
August 14, 2026
Latest release
v2.14.0 · August 11, 2026

Figures as of . Project site: tt-a1i.github.io.

Ask an agent for an architecture diagram and you get something that looks right. Whether the arrows correspond to calls that exist is a separate question, and usually the answer is partly. Archify is designed around that gap.

What it is

An MIT-licensed agent skill for Claude Code, Codex CLI, Cursor, OpenCode and Raven that turns a repository or a system description into an interactive technical map: five diagram types (architecture, workflow, sequence, data flow and lifecycle) with four visual presets, dark and light themes and optional finite motion.

Output is one self-contained HTML file, plus PNG, SVG, WebM and a 1200×630 share card.

Why it showed up now

v2.14.0 on August 11, four months after the first commit, 13,000 stars. The project publishes a gallery of output it calls the Proof Lab.

How it actually works

Between the agent and the picture sits a typed JSON intermediate representation with deterministic checks. The model proposes topology; the IR validates it; the renderer draws only what survives. This is why the project can use the word verifiable about output that a language model helped produce.

The consequences show up in the interactions. You can search nodes, optionally open revision-verified source for a node, trace upstream and downstream reach along exact routes, compare roles, and play guided stories, and none of these lets the diagram invent topology to fill a gap.

The feature I would use first is the snapshot diff: compare two validated snapshots as Before / Delta / After and get exact added, removed, changed, moved and rerouted facts. Reviewing an architectural change as a diff of validated graphs beats reading two diagrams side by side.

Try it

npx skills add tt-a1i/archify -g

Then ask your agent to map the repository's runtime architecture. The Proof Lab shows generated output before you install anything.

Where it is weak

Verifiable means the graph is internally consistent and node references check out. It does not mean the abstraction is the right one. Deciding which components matter and which detail to omit is the hard part of architecture drawing, and that judgement still comes from a model.

Self-contained HTML with motion and an interactive graph is a heavy artefact. For a large system, that single file can be slow to load.

Fifteen open issues, v2.14.0 in four months. The README carries two sponsors, one of them an API reseller with a referral link and a discount code, the other a company whose own agent harness supports Archify as a skill. Both are disclosed; read the README with that in mind.

Sources

  1. tt-a1i/archify · GitHub · April 15, 2026
  2. Release v2.14.0 · GitHub · August 11, 2026