Lavish (lavish-axi) and the AXI principles

Maker: Kun Chen (@kunchenguid)
URL: https://github.com/kunchenguid/lavish-axi ; https://axi.md (repo https://github.com/kunchenguid/axi)
Status (10 October 2026): active open-source project. MIT, JavaScript (TypeScript-checked), about 4,050 stars and 370 forks, created 11 May 2026, last push 10 October 2026. Released through release-please: v0.1.85 on 9 October 2026, v0.1.84 on 7 October, v0.1.83 on 6 October, so roughly a release a day. GitHub’s Releases tab is empty; the CHANGELOG is the record.

What it is

“HTML is the new markdown. Lavish is the new editor for your HTML artifacts.” A local CLI plus a browser editor for HTML files that an agent writes. The agent opens a file in a review session; the person reads it, annotates elements and text, edits Mermaid diagrams as whiteboards, attaches images and queues notes; the agent collects the feedback with a long poll and revises the file, which live-reloads.

The problem it’s solving

Agent-written HTML is a one-way broadcast. The person sees a page but has no precise way to point at a part of it and say “this is wrong”, and the agent has no cheap way to wait for that. Lavish closes the loop: targeted feedback in, cheap waiting out.

Its path / bet

Accept HTML as the agent’s output format, and build the review loop around it, locally, as a tool any agent can drive from a shell. It’s an “AXI”, Chen’s term for an Agent eXperience Interface: a CLI designed for agents first.

How it works

lavish-axi <file.html> starts (or reuses) a detached local server (port 4387 by default) and opens a review session keyed by file path. The artifact runs in a sandboxed iframe; Lavish injects a small SDK for annotations, snapshots and layout checks without changing the saved file.
lavish-axi poll <file> long-polls until feedback arrives, the session ends or every reviewer disconnects. Feedback carries the selected text, range anchors and CSS locators (table clicks include row and column names); images arrive as local file paths. Per the changelog, poll output also carries a DOM snapshot and browser layout warnings.
reply posts an agent reply (a Markdown subset) into the chat panel; poll --agent-reply replies and polls again. Delivery receipts show whether a note was seen, worked on or answered.
export makes a standalone copy; share publishes to a third-party host (ht-ml.app), with password-protected private links since 0.1.60; Tailscale binding lets a phone on the tailnet open a session.
“Playbooks” and a design skill ship with the CLI to steer how the agent writes pages (code walkthroughs, explanations of existing systems, and so on).
Installs as an agent skill (npx skills add kunchenguid/lavish-axi --skill lavish) or with hooks.

AXI’s ten principles (axi.md): token-efficient output (TOON, claimed about 40% fewer tokens than JSON); minimal default schemas; truncation with a size hint and --full; pre-computed aggregates; definitive empty states; structured errors, idempotent mutations, no interactive prompts; ambient context through session hooks; content first (a bare command shows live data); contextual help[] next steps; concise per-command --help. The site reports benchmarks (Claude Sonnet 4.6 as agent and judge; 490 browser runs and 425 GitHub runs) in which AXI tools matched or beat MCP equivalents at a fraction of the cost (for example $0.050 versus $0.101 to $0.148 per GitHub task). The author notes the limits: public sites, one model, read-only tasks, an LLM judge.

Strengths

The loop is right: agent writes, person points at the exact thing, agent waits cheaply, agent revises. It’s the nearest thing in the browser to what fictty wants in the terminal.
Precise, structured feedback (locators, ranges, row and column names) rather than “the third chart looks off”.
Local-first, no account, any agent that can run a shell command.
Layout warnings reported back to the agent: a partial answer to “the agent can’t see what it drew”.
Shipping daily, with a clear voice and a growing user base.

Weaknesses / limits

The artifact is still a static file the model writes. Data in it is frozen at write time unless the agent rewrites the file; there is no binding to live sources.
Read-back is a DOM snapshot plus annotations, not the frame the person sees.
Needs a browser and Node. Over SSH it relies on Tailscale or a share link.
Every revision is a regeneration of HTML; long pages cost output tokens and drift.
Single maintainer; the format and commands are young and moving.

Relation to fictty

Inspire, edging into compete. It solves the same loop for the same users (people supervising a coding agent), in the browser. If people are happy to review in a browser tab, Lavish plus Claude Code Artifacts covers much of fictty’s “review queue” and “walkthrough” cases. fictty’s distinct ground is live data bound without the model, keyboard-speed interaction, exact frame read-back, and running in the pane over SSH.

Could fictty adopt it instead of building?

Not for fictty’s core: it’s a browser HTML editor, not a runtime. But for the “this screen became a document” path, yes: rather than build HTML review, fictty should be able to export a screen to a self-contained HTML file and hand it to Lavish or an Artifact. Interop costs little; rebuilding annotation and whiteboards would be a waste.

What fictty should take from it

watch as a long poll that returns structured feedback, with the same three exits Lavish has (response, session ended, everyone gone) and an “ended” state that tells the agent to stop. This moves watch to the top of the plan, as the first landscape pass already said.
Point-at-it feedback. Let the person mark a node, row or cell and attach a note; deliver it with the node id and row key. fictty can be more exact than CSS locators because every node has an id.
Delivery receipts (seen, working, answered) shown in the screen itself.
Layout warnings to the agent: fictty knows when text is truncated, a column is squeezed or a node is off-screen. Report it in get --stats or a lint verb.
Follow AXI for the CLI: content-first bare commands, definitive empty states, help[] next steps, compact output. Consider TOON or a compact text form for get --state.
Playbooks: ship recipes and guidance as an agent skill, installed the same way.

Sources

kunchenguid/lavish-axi, its CHANGELOG and the GitHub API record (fetched 10 October 2026)
Kun Chen, AXI: Agent eXperience Interface and kunchenguid/axi

Not verified: the AXI benchmark numbers (self-reported, single model), and the exact contents of the DOM snapshot in poll output (inferred from changelog entries, not tested).