GitHub repo explainer · local artifact collaboration

lavish-axi

Lavish AXI is a local browser review loop for agent-generated HTML artifacts. An agent writes a page, opens it through lavish-axi, the human annotates exact elements/text in the browser, and the agent long-polls those comments back into the coding session.

Source: github.com/kunchenguid/lavish-axi npm: lavish-axi@0.1.27 Stack: JavaScript / Express / browser client License: MIT Repo signal: 376★ · 29 forks · 9 open issues

What it is

Lavish turns static HTML deliverables into a feedback surface. It does not replace the artifact; it wraps it in a local editor chrome, injects a small SDK, records targeted annotations, then hands those prompts back to the agent via CLI polling.

Agent-native: installed as an Agent Skill or invoked directly through npx -y lavish-axi.
Local-first: artifact files and session state stay on the machine by default.
HTML-first: built for plans, diagrams, comparisons, reports, prototypes, and other rich artifacts where screenshots + prose feedback are lossy.

Why Ananth should care

Mission Control already uses rich HTML pages for research, planning, and options artifacts. Lavish is directly aligned with the next bottleneck: Ananth reviewing those pages without needing to explain “the third card under the roadmap is wrong.”

Could make Dab → Claude Code planning pages reviewable on a visible browser.
Could reduce token-heavy back-and-forth by turning visual edits into targeted prompts.
Especially relevant for Notice Board artifact pages, product specs, UI explorations, and comparison tables.

Core loop

1. Agent writes
A standalone artifact.html file.

→

2. Lavish opens
lavish-axi <file> serves it in a local browser UI.

→

3. Human annotates
Select text, click elements, queue comments.

npx skills add kunchenguid/lavish-axi --skill lavish
/lavish let's discuss our plan here

# direct path, no skill install needed
npx -y lavish-axi path/to/artifact.html
npx -y lavish-axi poll path/to/artifact.html

Best use cases

Technical/product plans
Option comparisons
Architecture diagrams
Research reports
UI/design explorations
Diff and review summaries

Not the point

It is not a general CMS.
It is not an auth-protected public review app.
It does not magically make every agent accept feedback unless that agent runs the poll loop or skill.

Fit score

8/10

Strong fit for Dab/Claude Code artifact review. Needs an experiment on the Pi/Mac/browser path before adopting as a default workflow.

Source map

src/cli.js

Main command router: open, poll, end, stop, server, playbook, design, setup. Normalizes lavish-axi artifact.html into an open command and manages spawned server/session behavior.

src/server.js

Express server, session creation, long-polling, browser/server event stream, artifact serving, SDK injection, file watching, and idle shutdown.

src/artifact-sdk.js

Injected artifact-side SDK used for annotations, DOM snapshots, action controls, queued prompts, and communication with the chrome wrapper.

src/chrome-client.js + src/chrome.css

Browser chrome UI for sessions: composer, annotation cards, queue/send flow, feedback display, reload/copy/end controls.

src/session-store.js

Local state store keyed by canonical file path, preserving sessions and queued feedback under the Lavish state directory.

src/playbooks.js / src/design-reference.js

Agent guidance for generating better artifact forms: diagrams, tables, comparisons, plans, diffs, inputs, and slides; includes design fallback advice.

skills/lavish/SKILL.md

Installable Agent Skills entrypoint. Tells Claude Code/Codex/OpenCode-style agents when to use Lavish and which commands to run.

Signals and maturity

Signal	Observed	Implication
Release state	Published package `0.1.27`; latest GitHub release on 2026-06-11.	Active, fast-moving early project. Expect churn.
CI / checks	README advertises CI and release workflows; package scripts include build, lint, format, typecheck, node tests, and skill build check.	Engineering hygiene is better than a throwaway demo.
Runtime	Node `>=22`, Express 5, chokidar, open, axi-sdk-js, Tailwind/DaisyUI design assets.	Works naturally in modern agent/dev machines, but Pi/older Node setups need version check.
Community signal	376 stars, 29 forks, open issues/PRs around remote SSH, Copilot/Antigravity behavior, targeted messaging.	Concept resonates; integration semantics across agents are still settling.
License	MIT.	Safe to experiment with and adapt inside Dab workflows.

Caveats

Local server exposure: default loopback is safe-ish, but remote/SSH workflows require conscious host binding. Open issues/PRs discuss non-loopback/Tailscale use and unauthenticated local server tradeoffs.
Agent compatibility: the cleanest path is agents that understand skills and can run lavish-axi poll. Some IDE agents may not resume automatically.
Artifact assets: local sibling assets are supported; root-prefixed paths are discouraged because Lavish serves the artifact through its own route.
Version churn: project was created in May 2026 and is iterating quickly. Adopt experimentally, not as a core dependency yet.

Minimal Dab experiment

Pick one non-sensitive Mission Control planning artifact or create a scratch /tmp/lavish-test.html.
Run npx -y lavish-axi --no-open /tmp/lavish-test.html on the machine where Claude Code is running.
Open the session URL in the Pi/Mac browser depending on where the server is reachable.
Run npx -y lavish-axi poll /tmp/lavish-test.html in the agent session.
Annotate 2–3 exact UI/text changes; verify they come back as actionable prompts.

Do not point it at sensitive Notice Board pages until the local/remote access model is understood.

Recommendation for Mission Control

Do not install it globally as a default yet. First run a focused experiment with one Claude Code-created HTML page. If the poll loop feels natural, Lavish can become the review layer for higher-touch artifacts: options pages, product plans, implementation reviews, and visual specs. It is less useful for simple link explainers like this one, because the Notice Board already provides enough read-only navigation.

Likely integration lane: Dab writes/asks Claude Code for an artifact → launches Lavish for visual review when Ananth is actively present → receives targeted feedback → final artifact is saved back into Mission Control. This should remain an interactive command/control workflow, not an autonomous cron behavior.