kunchenguid/lavish-axi

HTML is the new markdown. Lavish is the new editor for your HTML artifacts.

Agents are good at producing rich HTML artifacts, but the human-agent collaboration loop on such artifacts is lacking and falls back into screenshots and long responses for “tell me what to change.” That loses the thing HTML is best at: interactivity.

Lavish Editor opens agent-generated HTML files in a local browser, lets you pinpoint elements and selected text, edit diagrams your agent authored as Mermaid whiteboards, and send feedback to the agent to address.

• Local-first - Review local HTML artifacts with a local CLI and no cloud dependency in the core feedback loop; hosted sharing through third-party ht-ml.app is explicit and opt-in.
• Human-AI collaboration - Annotate elements and selected text ranges, edit Mermaid whiteboard diagrams, and send messages to the agent without leaving Lavish Editor.
• Battery included - Lavish Editor teaches your agent good visualization for common use cases such as product or technical plans, design explorations and more out of the box.

Lavish Editor is an AXI, which means -

• It's just a CLI any capable agent can run without setup.
• It's optimized for agent ergonomics. TOON output, long polling, and contextual disclosure making it highly token efficient.
• The skill and hooks below only handle discovery; agents learn to use the AXI by using it.

Quick Start

Install the Lavish skill in the Agent Skills format with npx skills:

npx skills add kunchenguid/lavish-axi --skill lavish

That is the entire setup - no npm install needed. The skill teaches your agent to run Lavish through npx -y lavish-axi, so the CLI comes along on demand. It stays a short stub and sends the agent to npx -y lavish-axi --help, design, and playbook for current instructions, so an installed copy cannot go stale against a newer CLI. Its frontmatter also includes Hermes Agent metadata, so Hermes-compatible harnesses can categorize and surface it as a first-class productivity skill. This installs the public lavish skill. The repository also contains an internal lavish-design brand skill for maintainers; default npx skills add ... --list and skills.sh discovery hide it unless INSTALL_INTERNAL_SKILLS=1 is set.

Then, in agents that expose skills as slash commands (Claude Code, for example), invoke it directly:

/lavish let's discuss our plan here

Or just ask for anything that is easier to grasp visually - a plan, comparison, diagram, table, code view, or report - and the agent loads the skill on its own when it recognizes the task.

By default the skill lands in the current project's skills directory (.claude/skills/, for example); add -g to install it for all projects (~/.claude/skills/).

Other Ways to Use Lavish

The skill is the recommended path, but it is not the only one.

Zero setup

Lavish is an AXI, so any capable agent can run the CLI directly with nothing installed at all. Just tell your agent:

Use `npx -y lavish-axi` to write a product or technical plan for what we discussed.

Session hook

Want Lavish's ambient context - including your live open sessions - fed into every agent session instead of loading on demand? Install the CLI globally and opt into the hook:

npm install -g lavish-axi
lavish-axi setup hooks

This installs a SessionStart hook for Claude Code, Codex, OpenCode, and GitHub Copilot CLI that surfaces open sessions, visualization playbooks, and usage guidance at the start of each session. Unlike the skill, the hook also shows your live open sessions, so a fresh agent session can resume an in-flight review. Restart your agent session after running this so the new hook takes effect.

Agent Plugin

Lavish also ships as an Agent Plugin - the vendor-neutral packaging standard for skills and MCP servers - so clients that speak that format can load it directly.

No marketplace is involved. The installed npm package is the plugin: plugin.json sits at the package root next to the skills/ directory, so whatever npm install already put on disk is a complete, conformant plugin. Install the CLI, then register it:

npm install -g lavish-axi
lavish-axi setup plugin

That registers the installed package with every supported client it finds - VS Code, Cursor, and GitHub Copilot CLI - and reports which ones were absent. It is opt-in and idempotent, and it repairs the registered path after a reinstall or relocation. Reload each client afterward.

Each client is registered independently: one that cannot be registered is reported with what to do about it, and never blocks the others or fails the command.

To register by hand instead, point any client at the package directory (npm root -g/lavish-axi):

Codex and ChatGPT install plugins only from marketplace sources, so Codex users should use the session hook above instead. Lavish declares no MCP server - the CLI itself is the agent interface - so a plugin install brings the same lavish skill, and the skill and plugin are alternatives rather than a stack.

From source

git clone https://github.com/kunchenguid/lavish-axi.git
cd lavish-axi
pnpm install --frozen-lockfile
pnpm run build
pnpm link

How It Works

┌───────────────┐
│ Agent writes  │
│ artifact.html │
└───────┬───────┘
        ▼
┌────────────────────────┐
│ lavish-axi <file_path> │
│ opens local browser UI │
└───────┬────────────────┘
        ▼
┌────────────────────────┐
│ Human annotates text   │
│ or elements, sends     │
│ chat, or queues layout │
│ issues from the inbox  │
└───────┬────────────────┘
        ▼
┌────────────────────────┐
│ lavish-axi poll waits  │
│ and returns prompts    │
│ the user queued        │
└────────────────────────┘

• File-path identity - Sessions are keyed by the canonical HTML file path, so agents do not need opaque IDs.
• Portable artifacts - The artifact runs in a sandboxed iframe while Lavish injects a small SDK for annotations, snapshots, feedback controls, and render-time layout checks. Author-defined links and popups can open in top-level tabs, while artifact documents remain sandboxed without same-origin access. Lavish does not inject any design system, so the saved HTML file renders identically whether you open it through lavish-axi or directly in a browser. Run lavish-axi design for the single source of agent-facing design guidance, including optional CDN snippets and the whiteboard (Mermaid) opt-in snippet.
• Self-paint warning - lavish-axi <html-file>, export, and share run a render-free check for artifacts missing an explicit page background and return a one-line self_paint_warning. The check fails open - any stylesheet link, @import, Tailwind runtime script, color-scheme, or html/body/:root background signal suppresses it - and it never blocks the open.
• Open-time layout gate - The browser chrome masks an artifact only while the real in-iframe audit waits for fonts and final geometry. The first completed client-side check reveals the artifact, whatever it found and even if reporting that check to the server fails; the gate never holds the review hostage waiting for a repair or a network round-trip. The user can click Show anyway, and a bounded safety timeout fails open from every gate state. If the review cannot load at all - the chrome's own script never runs, or the server does not answer the artifact's load request after several retries - the mask names the problem and offers Check and reload without removing the independent Show anyway escape. A review already loaded in another browser tab is named the same way, with a Take over here button that moves it into the current tab, because Lavish loads an artifact in one tab at a time.
• Layout issues inbox - Detection is passive. After fonts and finite animations settle, the injected SDK confirms severe failures from direct rendered evidence such as materially escaped meaningful content or required controls, clipped text fragments, viewport reachability, or near-total semantic occlusion. Explicit ellipsis and line clamp, standard visually hidden accessibility text, intentional scrollers or masks, parent overhang, generic element scroll geometry, decorative overlap, and uncertain motion do not produce findings by themselves. Proven failures are filed in a Layout issues button in the top bar, which is hidden while nothing is unresolved and otherwise shows the unresolved count. Its drawer lists each issue with severity, a plain-language explanation, the affected viewport, the target/component identity, when it was last seen, and its lifecycle state, plus per-issue Reveal (highlight it in the artifact) and Dismiss actions. Nothing is selected by default. The user picks issues (or Select all) and Queue selected fixes turns that whole group into one ordinary queued prompt, tagged layout-warnings, that reaches the agent through the normal feedback path when they send. Detection never returns lavish-axi poll and never wakes an agent; only the user queueing a fix does. The one exception is a fatal artifact_failures response, for failures that make the review itself unusable, such as the artifact document or one of its own local assets failing to load.
• Layout issue lifecycle - Each issue is identified by a stable fingerprint of the diagnostic rule, the normalized target identity, and the viewport class, so repeat detections update one record instead of inflating the count. Open means the latest completed check for its viewport still detects it. Queued for fix means the user asked for a repair - it stays unresolved and counted, and cannot be queued again while that request is outstanding. Resolved requires a newer successful artifact load plus a complete check at the same viewport that no longer detects it; it then leaves the count but keeps a bounded history. Still present (recurring) means a queued issue survived a newer revision, so it is selectable again with its earlier attempt retained. Unverified means a reload or check failed or was incomplete, so the prior issue was preserved rather than cleared. Returned means a resolved issue came back on a later revision. Dismissal applies only to the current artifact revision; a later revision surfaces the issue again if it is still detected. A check at one viewport never clears an issue found at another, and a viewport removed from the configured diagnostic set (LAVISH_AXI_DIAGNOSTIC_VIEWPORTS, default all) is marked obsolete with an explicit reason rather than reading as fixed.
• Local assets - Copy local images, CSS, fonts, and scripts next to the HTML artifact and reference them with relative paths from that directory; root-prefixed paths such as /assets/logo.png will not resolve through Lavish's artifact route.
• Export and sharing - lavish-axi export writes <name>.export.html by inlining local assets only, stripping the annotation SDK, and leaving remote CDN/font references as links that still need network access. lavish-axi share publishes the same local-inlined HTML to ht-ml.app, a third-party hosting service not part of Lavish. Publishing sends the artifact to ht-ml.app's servers, public by default, or private and password-protected with --private (Lavish generates the password and returns it once, in the command output and in the browser publish dialog) or --password <pw> (one you supply, never echoed back). A generated password is a shared secret: give it to whoever should read the page, and expect it to appear in your agent's transcript, because the agent has to relay it. Lavish does not store it. The response includes a secret update_key shown once. Keep it to republish the same URL later with --site <site_id> --update-key <key>, which replaces the HTML in place and leaves the password alone unless you pass --private (rotate to a new generated one) or --password <pw> (set one). Locking a page that was public is not instant - it can keep answering from ht-ml.app's CDN cache for minutes after the password is set - while a page that was already private has no such cached copy. A page's password cannot be removed: ht-ml.app accepts a request to clear one and silently ignores it, so Lavish offers no way to make a private page public rather than telling you it did something the host did not do. Republish as a new page if you need a public URL. ht-ml.app has no delete endpoint, so --unpublish --site <site_id> --update-key <key> replaces the page with a short placeholder and locks it behind a discarded random password rather than removing it: the URL still resolves and the host still holds what was published. The content swap is immediate, but the lock is not: a page that was public can keep serving the placeholder to uncredentialed visitors from ht-ml.app's CDN cache for minutes afterwards. The update_key plus --private republishes real content behind a new password you can share. Bundling never fetches remote URLs, Lavish itself does not set a CSP, local reads stay confined and size-capped, and absolute file:// paths outside safe inlined asset references are redacted before output. Per-asset and per-bundle inline caps default to 10 MB and 25 MB, overridable with LAVISH_AXI_EXPORT_MAX_ASSET_BYTES and LAVISH_AXI_EXPORT_MAX_BUNDLE_BYTES. Unresolved local assets or export notices such as author-set CSP meta tags and redacted file URLs are surfaced in command or browser output. Use --token or LAVISH_AXI_HTML_APP_TOKEN for an optional bearer token when publishing a new page (a republish or --unpublish authorizes with the update_key instead and rejects --token); set LAVISH_AXI_HTML_APP_API_URL to override the ht-ml.app API base and point share at a backend you control (see Self-hosting the share backend for the contract it must implement).
• Live reload - Lavish watches the HTML artifact file by default and preserves review context across reloads: the artifact iframe scroll position, an open annotation card's unsent text, and answers to data-lavish-question controls (application-owned form state is left alone). Unsent annotation text also survives a full reload of the review page itself. While a queued layout-issue batch is outstanding, closely spaced saves coalesce so one batch of fixes costs one refresh. To also reload on sibling asset changes, add data-lavish-live-reload-root to the root element or <meta name="lavish-live-reload" content="root">. If the element an unsent annotation was attached to is gone from the artifact for good, Lavish cannot reopen that card, so it writes your text into the conversation panel under Unsent annotation - selectable, never written over anything you have typed, and kept there across reloads; no note is ever dropped to make room for a newer one, and a note the browser refuses to store says so where it is shown.
• Feedback controls - Native controls (radios, checkboxes, inputs, selects, buttons, labels, disclosure summaries, contenteditable) are interactive automatically, so they do not need data-lavish-action. For reversible choices, let option clicks update local state, then queue exactly one final answer from a per-question submit or Queue answer button with window.lavish.queuePrompt(). Mark only custom (non-native) clickable elements with data-lavish-action so Lavish does not annotate them, and use data-lavish-question or queueKey when pre-send updates for the same question should replace each other. On wider screens, queued annotation preview pills and chat history share a scrollable Conversation panel above a sticky composer, so long feedback queues do not push the text box or send controls off screen. The browser chrome keeps editing actions in the overflow menu (copy path, reload artifact, copy DOM snapshot, export standalone HTML, publish link, end session), while the composer exposes Send & End beside Send to Agent to submit queued prompts and user-ended attribution together.
• Reviewing on a phone - Below 860px wide, the artifact takes the whole screen above a Conversation dock, and the conversation opens as a bottom sheet over it: tap the dock, swipe it up, or press the chevron to raise it; tap the dimmed artifact, swipe the sheet down, press the chevron, or press Escape to lower it. The dock reports what matters while the sheet is down - how many prompts are queued, a reply that arrived while you were reading, or whether the agent is listening - and the sheet stays open across a reload of the review page. The sheet sizes itself to the visible viewport and respects safe-area insets; if the keyboard or attachments leave little room, conversation content yields or scrolls while the send actions remain pinned above the bottom edge. In landscape the sheet covers the top bar as well. Wider screens keep the side-by-side layout.
• Keyboard shortcuts - In the chrome composer, Enter sends queued prompts and Shift+Enter inserts a newline. In the annotation card, Enter queues the annotation, Shift+Enter inserts a newline, and Ctrl+Enter (Cmd+Enter on macOS) queues it and sends all queued prompts immediately. Escape closes the card, same as Cancel, but only while it is empty (no text, no attachment); with unsent text or an attachment present, Escape does nothing rather than risk discarding it. Cmd+I or Ctrl+I toggles between annotate and explore mode from either the browser chrome or the artifact iframe, including while focus is in a textarea or control.
• Agent presence - The browser shows when no agent is listening, keeps queued feedback for the next successful lavish-axi poll send even across reloads, and keeps human feedback actions available while the agent is working because the server queues them for the next poll. The agent's reply (--agent-reply) concludes delivered work and returns presence to waiting. The no-timeout poll always writes an immediate stderr banner so it is visibly not hung; it adds the periodic stderr wait ticks only in an interactive terminal, so when stderr is piped (as under agent harnesses) the captured output carries no tick noise. Stdout always stays reserved for the final response; if the poll is interrupted or times out before feedback arrives, re-run it because feedback remains queued until delivery. Poll delivery consumes the response, so read the complete response before truncating or filtering it. Codex-specific guidance keeps that poll attached to the active turn instead of hiding it in a background task, because completed background tasks may not resume the agent.
• Session end etiquette - Lavish tracks who ended a session: a human clicking End session (or Send & end session) in the browser is a user-initiated end, while lavish-axi end <html-file> is agent-initiated. When either side ends the session, every open review tab becomes visibly read-only and disables its feedback controls; feedback submitted after the end is refused instead of being accepted without an agent to receive it. A plain lavish-axi <html-file> after a user-initiated end refuses to reopen the browser and returns guidance instead; pass --reopen only when the user asks for further review or something important needs their visual attention. Agent-initiated ends keep reopening normally, same as before. lavish-axi poll's ended response and the feedback response for the final batch before an end both carry next_step guidance telling the agent to stop polling and deliver remaining updates in chat instead of reopening.
• Precise targets - Text annotations include selected text plus range anchors, and text selections carry those anchors only. Clicking an element inside a table also carries the cell's visible row and column names alongside the exact CSS locator, so filtered or sorted rows do not make feedback look misdirected. When merged cells make either name ambiguous, Lavish leaves that name out rather than guessing; an explicit <th scope="row"> remains authoritative even when a rowspan makes the row's position ambiguous. The CSS locator still points at the exact element you clicked, so an annotation with an omitted name is only less descriptive, never mislabelled.
• Image attachments - Attach reference images (PNG, JPEG, WebP) in either the Conversation composer or an annotation card by pasting, drag-dropping, or using its image picker; each image shows a thumbnail chip with upload, remove, retry, and error states. A paste containing both text and images preserves the text while attaching the images (a copied file's own name or path is treated as placeholder and not pasted). Conversation messages may contain text and images together or images only. Images are stored under the state dir and the queued prompt carries a server-generated absolute path and content-hash id (plus mime and dimensions) - never the raw bytes - so lavish-axi poll hands the agent a local file path to open. Limits are LAVISH_AXI_MAX_ATTACHMENT_BYTES (default 10 MiB per image), LAVISH_AXI_MAX_ATTACHMENTS_PER_PROMPT (default 4), and LAVISH_AXI_MAX_PROMPT_ATTACHMENT_BYTES (default 25 MiB per prompt); if any image is missing or any prompt breaches a count or byte cap, the entire send batch is rejected, the queue is preserved, and the reason is surfaced in the composer rather than silently dropping images. As a browser-side abuse guard, each chrome page allows 30 upload attempts per rolling minute, 4 uploads in flight at once, and 256 MiB of attempted image bytes over its lifetime; rejected uploads stay visible for retry or removal. Attachments are cleaned up by LAVISH_AXI_ATTACHMENT_TTL_MS (default 7 days; 0/off disables) but only once no pending prompt still references them; LAVISH_AXI_MAX_ATTACHMENT_DISK_MB (default 512 MiB; 0/off disables) caps total attachment disk. The cap is enforced when an image is uploaded, not just periodically: the upload first reclaims unreferenced files (oldest first, and never one added within the last hour), and if it still would not fit, that upload is refused with a storage-full error on its chip instead of discarding an image the user is about to send.
• Mermaid diagrams - Whiteboards are an opt-in: agents author a diagram as Mermaid only when you ask for an editable whiteboard, and hand-authored inline SVG illustrations are the default figure medium otherwise. In the Lavish browser, every rendered Mermaid diagram in a .mermaid container becomes an embedded editable Excalidraw whiteboard. Click a diagram to unlock editing, and use its Fullscreen action to edit it over the whole viewport. Whiteboard scenes autosave locally. If a live reload changes the Mermaid source, an unmodified whiteboard silently re-converts to the new diagram. If the reviewer had edited the scene, reopening it lets them re-convert and discard the saved edits or keep editing the saved scene. Use Queue feedback to add a bounded edit summary plus local .excalidraw scene and PNG preview paths to the Conversation panel, then click Send to Agent to deliver it. The agent updates the artifact's Mermaid source, which remains authoritative. Flowchart, sequence, class, ER, and state diagrams convert to editable shapes; other diagram types are images that reviewers can draw and annotate. Lavish changes only the browser view, so saved, standalone, and exported artifacts still render plain Mermaid.
• Server cleanup - The detached server stops after the last session ends when nothing is connected, or after LAVISH_AXI_IDLE_TIMEOUT_MS (default 30 minutes) with no browser or poll connections. Set LAVISH_AXI_IDLE_TIMEOUT_MS=0 or off to disable idle self-shutdown.
• Server upgrades - One background server serves every session, so upgrading lavish-axi while reviews are open makes the next lavish-axi <html-file> replace that server. Only the review page for the artifact being opened reloads itself once the replacement answers - and not even that one while you have unsent annotation text open, which gets the same banner instead so the reload is yours to make. Every other open review page keeps working and shows a banner reading "Lavish was updated. This page is running the previous version.", with a Check and reload button and a Dismiss button, so no page you are reading reloads on its own. After lavish-axi stop those pages say Lavish was stopped and to reload after you start it again, and a restart that only picks up a local build says that rather than claiming an update. Every Check and reload control asks the server whether it is running before it navigates; while nothing answers, the page stays where it is and says so, and a check that gets no answer at all says that instead of guessing. In-flight lavish-axi poll commands end with an interrupted-poll error and are safe to re-run; queued feedback is never lost, and annotation text you have typed but not queued yet survives the reload as described under Live reload. A page waits for the replacement rather than reloading into a port nothing is listening on, and tells the user to restart Lavish if it never returns.
• Local-first state - Session state stays under ~/.lavish-axi/ by default, or LAVISH_AXI_STATE_DIR when set.
• Diagnostic viewports - LAVISH_AXI_DIAGNOSTIC_VIEWPORTS sets which viewport classes the layout-issue inbox tracks (mobile, compact, desktop; comma-separated, default all). Warnings whose class leaves the set are marked obsolete with an explicit reason instead of silently reading as fixed.
• Server port - Set LAVISH_AXI_PORT to choose the server port; it defaults to 4387.
• Network binding - With Tailscale running, the review server automatically listens only on loopback (127.0.0.1) and this machine's Tailscale IPv4 address - never on 0.0.0.0 or every interface. The generated session link uses the machine's MagicDNS name, so it is the phone-ready URL to open from another device on the same tailnet. When Tailscale is absent or down, Lavish silently falls back to loopback-only and prints no phone URL. If Tailscale is running but MagicDNS is unavailable or its address cannot be bound after brief retries, Lavish visibly warns that phone access is unavailable and remains loopback-only. Binding beyond loopback exposes an unauthenticated server that can read and serve arbitrary local files to devices that can reach it, so the tailnet should be trusted. Any explicit LAVISH_AXI_HOST overrides automatic Tailscale binding; wildcard values are restricted to loopback, while a non-wildcard value selects that one concrete bind address. LAVISH_AXI_LINK_HOST controls the link host when automatic Tailscale binding is disabled.
• Allowed hosts - To defend against DNS rebinding, the server rejects (403) any request whose Host header is missing or not one it answers to: loopback names plus the concrete Tailscale IPv4 address and MagicDNS name when the Tailscale listener is successfully bound. If you configure a reverse proxy or another intentional hostname, list it in LAVISH_AXI_ALLOWED_HOSTS (whitespace-separated). Behind a reverse proxy, the forwarded X-Forwarded-Host is validated against the same list, so add the public hostname there and have the proxy send it together with X-Forwarded-Proto. Set LAVISH_AXI_ALLOWED_HOSTS to * to disable the check entirely, only when the server sits behind your own authentication or proxy. Mutating routes also reject a present foreign Origin or Referer (403); header-less CLI control requests remain allowed where supported.
• Browser opening - Set LAVISH_AXI_NO_OPEN=1, equivalent to --no-open, to create or resume a session without launching a browser window.

CLI Reference

Known playbook IDs: diagram, table, comparison, plan, code, input, slides. One artifact often combines several playbooks, such as a plan that includes a comparison and a diagram, so agents must match against each use_when trigger and open every matching playbook before writing HTML. For flows, architecture, state, or sequence diagrams, open the diagram playbook for the recommended tooling and SVG guidance.

Flags

Development

pnpm run check          # Run all verification commands
pnpm run build          # Bundle the publishable CLI, chrome, and design assets
pnpm run build:skill    # Regenerate the installable lavish skill
pnpm test               # Run node:test tests
pnpm run lint           # Run ESLint
pnpm run format:check   # Check Prettier formatting
pnpm run typecheck      # Run TypeScript checkJs validation