lavish-axi

lavish-axi is an open-source CLI tool built by Kun Chen that bridges the gap between AI-generated HTML artifacts and human review. It opens agent-produced HTML files in a local browser, lets users pinpoint elements or selected text, and sends structured feedback back to the agent — all without leaving the editor. The project is published on npm under the MIT license and has accumulated over 1,200 GitHub stars since its creation in May 2026.

What It Is

lavish-axi is an AXI (Agent eXtensible Interface) — a CLI that any capable AI agent can invoke directly without prior setup. It solves a specific problem in human-AI collaboration: agents are good at producing rich HTML artifacts (plans, comparisons, diagrams, reports), but the feedback loop typically degrades into screenshots and long text descriptions. lavish-axi restores interactivity by letting humans annotate the live rendered artifact and return precise, structured prompts to the agent.

How the Collaboration Loop Works

The core workflow follows four steps:

• The agent writes an HTML artifact to a local file path.
• lavish-axi <file_path> opens a local browser UI with the artifact in an iframe.
• The human annotates text ranges or DOM elements, types feedback, or interacts with native controls.
• lavish-axi poll long-polls and returns the human's prompts to the agent, including any layout warnings detected at render time.

Sessions are keyed by canonical file path, so agents never need opaque session IDs. The server binds to loopback by default, keeping the entire feedback loop local with no cloud dependency unless the user explicitly runs lavish-axi share.

AXI Architecture and Agent Ergonomics

lavish-axi is designed around agent ergonomics rather than human developer ergonomics. Key design choices include:

• TOON output (Token-Optimized Output Notation), long polling, and contextual disclosure to minimize token usage.
• Skill-first discovery — the recommended setup is a single npx skills add command that installs a skill file in the Agent Skills format; the CLI itself is fetched on demand via npx -y lavish-axi.
• Session hooks for Claude Code, Codex, OpenCode, and GitHub Copilot CLI that surface open sessions and visualization playbooks at the start of each agent session.
• Playbook system — built-in playbooks (diagram, table, comparison, plan, code, input, slides) teach agents good visualization patterns for common artifact types.

Layout Safety and Feedback Controls

lavish-axi includes an open-time layout gate: the browser chrome masks each artifact until an in-iframe layout audit reports no error-severity findings. The audit checks for horizontal overflow, element overflow, clipped text, and overlapping text, and returns findings via lavish-axi poll as layout_warnings with selector, kind, overflow measurements, viewport width, and severity. This lets agents fix unreadable layouts before asking the human to review.

Feedback controls support native HTML inputs automatically. Custom clickable elements can be marked with data-lavish-action. The window.lavish.queuePrompt() API lets artifacts queue structured answers from interactive controls before a single submit.

Export, Sharing, and Local-First State

• lavish-axi export writes a portable single-file HTML with local assets inlined and the annotation SDK stripped.
• lavish-axi share publishes the inlined artifact to ht-ml.app, public by default, with optional password protection; the response includes a one-time secret update key.
• All session state is stored under ~/.lavish-axi/ (or LAVISH_AXI_STATE_DIR) with no cloud dependency in the core loop.
• Per-asset and per-bundle inline caps default to 10 MB and 25 MB respectively, configurable via environment variables.

Update: v0.1.32

The latest release is lavish-axi v0.1.32, published on June 27, 2026. The repository shows active development with CI and release-please automation, 91 forks, and 21 open issues as of early July 2026. The project is written primarily in JavaScript and distributed via npm.