Render agent output as self-contained HTML. Ask for an artifact in plain language; the model composes a structured spec and a deterministic core assembles one self-contained, sanitized, themed file and opens it. The core knows no host tool; the first adapter is a Claude Code plugin.
/plugin marketplace add ibrahemid/to-html
/plugin install to-html@ibrahemid
Just ask, in plain language:
- "build me html of what we just did"
- "put this in a status html"
- "show me options for X"
- "make a checklist to test this"
The model picks a kind, composes the content, and opens the file. Kinds:
| kind | for |
|---|---|
dashboard |
status / handoff: sections of items with status, detail, links, copy-able prompts |
report |
tables + links (supports a bare plain table) |
options |
2 to 4 option cards to compare and pick |
diagram |
a system/flow diagram from nodes and edges |
checklist |
groups of checkable items (state saved in your browser) |
asset-grid |
a grid of downloadable assets |
findings |
an audit/inventory list with severities |
Every artifact is one self-contained file (inline CSS/JS/SVG, no external assets), sanitized, and theme-aware. Configure the default look and the browser it opens in:
/to-html config theme dark
/to-html config opener Dia
The model authors a structured spec, not raw HTML; a deterministic, zero-dependency core assembles it into the final file. It does not touch your normal replies: no auto-render, no formatting contract.
core/- the renderer and the artifact kinds. Zero runtime deps, deterministic.shared/- tool-agnostic transcript parsing.adapters/claude-code/- the Claude Code plugin (skill + CLI).cli/- the standalone CLI.scripts/- bundle sync and version propagation.
MIT.