kunchenguid/lavish-axi

What it does

Lavish closes the review loop on the HTML artifacts coding agents produce. Instead of describing changes in long chat messages or screenshots, you open the generated file in a local browser UI, annotate elements or text ranges, edit Mermaid whiteboards, and queue feedback that the agent retrieves through a long-polling CLI. Sessions are keyed by the canonical HTML file path, so agents do not need opaque IDs, and the artifact runs in a sandboxed iframe with an injected SDK for annotations, snapshots, and render-time layout checks.

The project targets people using capable coding agents (Claude Code, Codex, OpenCode, GitHub Copilot CLI, and others) who want a richer collaboration surface than a chat transcript. It also provides layout-issue detection: after fonts and animations settle, severe rendering problems are collected into an inbox the user can selectively queue as a fix prompt. Detection is passive and never wakes an agent by itself.

Quick start

The recommended path installs only a skill; the CLI is fetched on demand via npx.

1. Install the skill with the Agent Skills CLI: npx skills add kunchenguid/lavish-axi --skill lavish (add -g to install for all projects instead of the current project).
2. In an agent that exposes skills as slash commands, invoke it, e.g. /lavish let's discuss our plan here, or simply ask for a visual artifact such as a plan, diagram, or comparison.
3. The agent writes an HTML file and runs lavish-axi <file>, which opens the local browser editor.
4. Annotate, queue feedback, and send; the agent picks it up through lavish-axi poll <file> and posts replies via lavish-axi reply.

Alternatively, a global install (npm install -g lavish-axi) plus lavish-axi setup hooks installs a SessionStart hook for Claude Code, Codex, OpenCode, and GitHub Copilot CLI; restart the agent session afterward. lavish-axi setup plugin registers the npm package as an Agent Plugin in VS Code, Cursor, and GitHub Copilot CLI. Building from source requires pnpm and runs pnpm install --frozen-lockfile, pnpm run build, and pnpm link.

Usage constraints

The core review loop is local, but sharing is not: lavish-axi share publishes to the third-party host ht-ml.app, public by default or password-protected with --private/--password. A generated password appears in command output and in the agent transcript, since the agent relays it, and Lavish does not store it. Password removal is not supported by the host, and --unpublish replaces the page with a locked placeholder rather than deleting it, so the URL still resolves. A page that was public may keep serving from the host's CDN cache for minutes after locking.

Local assets should sit next to the artifact and use relative paths; root-prefixed paths will not resolve. export inlines local assets only and leaves remote CDN and font references as network links, with per-asset and per-bundle inline caps defaulting to 10 MB and 25 MB. The npm release sends anonymous usage telemetry to an Umami server by default; disable it with LAVISH_AXI_TELEMETRY=0. The project is MIT-licensed and lists macOS, Linux, and Windows in its badges.

HotRepo's take

The JavaScript daily list on 2026-11-11 recorded 50 new stars for the day and rank 2 on that language chart, against 4,532 stars and 418 forks overall. That single snapshot is a strong day on one language list, but it says nothing about sustained adoption; with no HotRepo daily all-language ranking recorded, the evidence is one dated spike rather than a trend. What stands out technically is the deliberate agent-facing design: long-poll delivery, typed listener and failure states, and a skill that stays a thin stub pointing at --help and playbooks so installed copies do not go stale. The telemetry-on-by-default and third-party sharing model are the parts a prospective user should weigh before adopting.

Project source and documentation