Project Description · Seed Document

Muse Stacks

A community website that tracks the best, most-used skills, connectors, and prompts for Muse — Meta's personal AI chat assistant — ranked by community popularity. Think ClawHub or Smithery, rebuilt for a consumer life assistant instead of a developer agent.

🔗 https://www.html-docs.com/site/muse-stacks/index

The MVP directory site. It may go live slightly after this document — if the link doesn't resolve yet, check back soon.

1Vision & problem

AI assistants keep adding "skills" (installable playbooks that teach the assistant how to do things) and "connectors" (OAuth integrations to apps like Gmail, Spotify, and Google Calendar). The catalog grows fast — and nobody can tell what's actually useful inside the assistant. Users scroll past long alphabetical lists, try things at random, and give up. Creators ship great playbooks and nobody finds them.

Muse Stacks is the community-ranked discovery layer for Muse. One trusted front page that answers "what should I add to my Muse?" the way Rotten Tomatoes answers "what should I watch?" — with popularity rankings, freshness signals, install counts, and use-case browsing, all driven by the community rather than by an app store team.

ClawHub (clawhub.ai) and Smithery (smithery.ai) solved this for developer agents; Muse Stacks solves it for a consumer life assistant — where the "app store" is the assistant itself, and installs should feel as easy as chatting.

2Research findings

Ecosystem research was completed 2026-09-15 and is summarized below. Claims marked verified live were read in a real browser that day; index-only claims come from web-index snippets and were not re-verified. The full report lives with the project notes.

The eight steal-worthy ideas

  1. Three-tab install panel: CLI / one-liner / "copy this prompt to your agent."
    ClawHub skill page — verified live at clawhub.ai/anotb/skills/homeassistant-skill. The Prompt tab is a pre-written message with safety-vetting instructions ("inspect metadata before installing…"), required env vars, required binaries, and the skill URL — designed to be pasted to your agent, which does the install.
  2. "Add to toolbox" card buttons + a personal Toolbox (cart).
    Smithery — verified live at smithery.ai. Cards carry one-click "Add to toolbox" buttons plus usage counts ("22.49k uses") as the headline social signal instead of stars. (Smithery was acquired by Arcade.dev in 2026 — also verified live via the site banner.)
  3. "Skills for the apps you already use" — app-tile browsing.
    ClawHub homepage — verified live at clawhub.ai. Zapier-style app-centric discovery (Gmail, Notion, Slack, WhatsApp…) grouped by Chat / Docs & specs / Web / Cloud. Directly fits Muse's connector concept.
  4. Organize by use case, price filter with a Free-mode slider, release-history feed.
    There's An AI For That — index-only: beehiiv case study. Organized by real use case ("remove background from 200 photos"), 1.7M newsletter subscribers with $0 paid ads.
  5. Programmatic SEO: pairwise pages (/skills/<a>/for/<use-case> or connector × skill) + app profile pages.
    index-only: Zapier's ~50k pairwise integration pages → 5–6M monthly organic visits (gracker.ai case study); TAAFT's machine-generated catalog + "claim this page" monetization ($49 basic, $347 "maximum exposure").
  6. In-chat OAuth connect link — authorize without leaving the conversation.
    index-only: Composio's composio_connect returns an in-chat OAuth link (github.com/k3-2o/pi-composio); Smithery's npx smithery mcp add notion → browser authorize URL flow (verified live at smithery.ai).
  7. Submit-your-skill self-serve: CLI publish + "import from GitHub" + versioning (latest tags, changelogs, rename/merge, pin installs).
    ClawHub — verified live homepage publish section at clawhub.ai.
  8. Freshness-as-trust: "re-checked daily, 90-days-quiet = delisted" + last-push dates in tables.
    index-only: pjdurden/claude-skills-live (GitHub) — anything archived, deleted, or 90 days quiet is flagged and delisted; hmzainjamil/awesome-claude-code's dead-link <7-day policy and tier emojis 🔥/⭐/🧪.

Other patterns worth noting

3Content model

Three entry types — kept deliberately simple, because a directory lives or dies on whether a new visitor instantly gets it:

Entry fields

The 8 categories

💰 Money✈️ Travel🏃 Health⚡ Productivity💬 Communication🛍️ Shopping📚 Learning🏠 Home

Categories are the shelf; use-case browsing is the front door. Visitors should be able to start from "plan my week", "find cheap flights", or "get fit" — à la There's An AI For That's use-case organization — and land on ranked entries, with categories as a secondary filter.

4Ranking approach

Seed-stage honesty: until real usage data exists, all scores are illustrative seed values — curated by the editorial team, never presented as download or install counts. The site labels them plainly as "seed ranking" until the first telemetry milestone replaces them.

5Install mechanics

Installing into a chat assistant should feel like chatting — not like configuring software. Four mechanisms, stacked:

  1. The 3-tab install panel (stolen from ClawHub). Every entry page shows three copy-buttoned tabs: Command (CLI one-liner), One-liner (npx-style, Vercel pattern), and Prompt for Muse — a pre-written setup message the user pastes to their assistant, which then performs the install. The prompt tab embeds safety-vetting instructions and required credentials, ClawHub-style.
  2. Agent-executable "Prompt for Muse" messages. The prompt isn't just instructions for a human — it's written so Muse itself can act on it mid-conversation: search the entry, fetch the playbook, and enable it. This is the Vercel find-skills pattern: the agent installs itself with zero context switch.
  3. 3-tap connector setup. For connectors: tap Connect → authorize in the app's OAuth screen → confirm in chat. The Composio pattern — the OAuth link arrives in-chat, so the user never leaves the conversation.
  4. "Add to Muse" toolbox cart (stolen from Smithery). A persistent cart: collect entries across the site, review the bundle, then install everything in one agent-executed prompt. The cart doubles as a shareable "skill stack" (see §6).

6Viral loops

7MVP scope

What shipped (or ships) in v1:

Deliberately out of v1: real usage telemetry, author dashboards, programmatic SEO pages, monetization. The directory earns trust first, scale second.

8Roadmap

  1. Real usage telemetry — install counts, 7-day retention, installer-only ratings; seed scores retire.
  2. Author dashboard — installs, errors, retention per entry (the Smithery publishing analytics model).
  3. Programmatic SEO pages/skills/<slug>, /category/<name>, /use-case/<slug>, plus pairwise connector × skill pages — the Zapier 50k-page playbook.
  4. Claim/verify monetization — the TAAFT model: creators claim their auto-generated pages and pay for verification tiers and newsletter placement.
  5. Deep-link install — if the Muse product ever supports it, one-tap muse:// (or universal-link) install cards in chat, mirroring the claude:// / cursor:// deep-link buttons used by MCP server authors.

9Open questions

  1. The true one-click surface. Does the Muse product expose a muse:// deep link or an agent-executed install command? Until validated with the product team, the "Prompt for Muse" copy-paste path is the install primitive.
  2. OAuth without leaving chat. How do connectors get authorized without breaking the conversation? The Composio in-chat link pattern is the target UX — but it needs product support for link rendering and token handoff.
  3. Moderation at scale. Manual security review works for 50 entries; it doesn't work for 5,000. What automated vetting (static analysis of SKILL.md, sandbox testing, permission diffing) gates the self-serve submit flow?
  4. External skill repos. Can Muse read external skill repos directly (GitHub raw, versioned bundles), or does every entry need to be mirrored/hosted by the directory? This decides whether Muse Stacks is a registry or an index.