archify vs Understand-Anything
Agent skill for architecture, workflow, sequence, data-flow and lifecycle diagrams: the agent emits typed JSON IR, Archify compiles it deterministically to interactive HTML/SVG with PNG/WebM export. — versus — Plugin for Claude Code and 16 other hosts that turns any codebase into an interactive knowledge graph — multi-agent analysis, layered dashboard, guided tours, diff-impact view, domain mapping.
understand-anything turns a codebase into an interactive knowledge graph inside Claude Code and 16 other hosts; Archify is the skill that turns a system description into a polished, exportable diagram. Overlap is 'a map in chat'; the source of truth differs.
| archify | Understand-Anything | |
|---|---|---|
| Stars | 49k | 81k |
| Forks | 3.2k | 6.8k |
| Language | JavaScript | TypeScript |
| License | MIT | MIT |
| Last activity | today | 2 days ago |
| Topics | skills, code-intel | code-intel |
| Curated connections | 3 | 6 |
archify — the curator's take
The right tool when an agent needs to *present* a system, not just describe it: `npx skills add tt-a1i/archify`, describe the system (or point at a repo) and you get a self-contained HTML map with search, upstream/downstream tracing, guided stories and a Before/Delta/After compare between two validated snapshots. The model never draws — it emits IR that a validator checks, so the topology can't be hallucinated at render time. NOT a code-derived architecture tool: what it shows is what the agent authored, verified against the schema, not against the code. For a map that is provably the codebase, use a code-intelligence tool and hand its output to Archify.
Understand-Anything — the curator's take
The onboarding killer app, and its motto is the right one: graphs that teach, not graphs that impress. /understand runs a multi-agent pipeline over the repo, then the dashboard gives you architecture layers, dependency-ordered guided tours, semantic search, diff blast-radius, and a business-domain view that maps code to real processes. The team trick is the sleeper: the graph is plain JSON — commit it once and every teammate skips the analysis. Incremental re-runs + a post-commit auto-update hook keep it fresh. Budget real tokens for the first full run on a large repo (their own warning), or point it at a local model. NOT an agent context server — this graph is for humans first; when you want a code graph served TO agents over MCP, that's codegraph-mcp or tokensave territory.