Tier List MCP

Create and rank tier lists inside an MCP-compatible chat. Name a category, drag options into S/A/B/C/D (or use the dropdowns), undo changes, and export a PNG. Chat-driven ranking also works.

Built with Manufact / mcp-use and MCP Apps. MIT licensed. Independent project; not affiliated with or connected to TierMaker.com. No TierMaker account needed.

The production build embeds the card JavaScript and CSS in its HTML resource, so sandboxed hosts do not need to fetch assets from localhost. UI changes are sent to the host through model-visible view state.

Use the hosted MCP

Install Tier List MCP — includes Cursor and VS Code install buttons and instructions for other clients. No local server or Manufact account needed to connect.

Paste this URL into your client's remote MCP settings:

https://tier-list.run.mcp-use.com/mcp

For Codex:

codex mcp add tier-list --url https://tier-list.run.mcp-use.com/mcp

Or add a Streamable HTTP server in Settings → MCP servers with the URL above, then save and restart.

For Claude Code:

claude mcp add --transport http tier-list https://tier-list.run.mcp-use.com/mcp

Then ask: “Make a tier list of AI coding tools.” Interactive cards require a client with MCP Apps support; other clients can use the tools and PNG previews. The shared endpoint is hosted on Manufact and is subject to hosting availability and usage limits. You can also run your own copy below.

Related MCP server: recraft-mcp-server

Run locally

Node 22.22.2 or newer:

git clone https://github.com/kbhuw/tier-list.git
cd tier-list
npm ci
npm run build
npm start -- --port 3000

Connect your MCP client to http://localhost:3000/mcp. npm run dev also provides the MCP Inspector. Remote clients need a reachable HTTPS deployment.

Deploy on Manufact

npx mcp-use login
npx mcp-use deploy --no-github

Use the endpoint returned by the deploy command in your MCP client. Alternatively import this repository in Manufact and use npm run build / npm start -- --host 0.0.0.0. Manufact deployment requires your own account. No paid plan or external API key is required by the application itself; hosting is subject to the platform's pricing.

Conversation

Make a tier list of agent frameworks.

The host agent researches or proposes candidates and calls create-tier-list. The card and PNG show every option unranked. Category discovery is performed by the host agent, using its own knowledge or search tools; this server does not contain an LLM or a hardcoded category catalog.

LangGraph in S, CrewAI in A.

rank-tier-list returns the complete revised board and a fresh image. Unmentioned options stay where they were.

Actually CrewAI in B. Put Mastra in A.

The next result shows the correction. To undo, the host passes the previous snapshot to show-tier-list.

Download.

download-tier-list returns PNG image content and an embedded binary resource. A file-capable host saves the resource's base64 bytes under the returned filename and attaches it. The MCP App supports drag-and-drop ranking, accessible tier dropdowns, Undo, and a Download PNG button; browser download behavior depends on the host's sandbox permissions. If blocked, the card displays the PNG and a Download in chat button. Download does not require opening a website.

Tools

Every tool returns structuredContent.board; retain this exact object for the next call. Tools are stateless: boards live in the conversation, so separate users never share mutable server state and server restarts do not lose a board. If a client discards the conversation's board, restore it from a saved JSON snapshot. There is no server-side account, saved-list browser, or cross-chat library.

Company cards include logos and names. Provide each option’s official source URL or logoDomain; creation fetches its favicon through Google’s fixed favicon service, normalizes it to a small PNG, and embeds it in the board. Pre-supplied PNG data URLs are also supported via logo. Embedded logos persist across moves and exports without further network requests. Missing or invalid icons fall back to the name. Optional descriptions and source URLs are preserved for the host. Arbitrary image URLs, browser credentials, and filesystem paths are not accepted. Limit: 80 options, 12 tiers. Inputs are validated; invalid moves fail the whole batch. Tool annotations are read-only because no server or external state is changed.

Evaluation

npm run build
npm run typecheck
npm test

Tests run a real HTTP MCP server on port 3418 and check tool discovery, view resources, all-unranked creation, incremental moves, correction, exact PNG resource export, and rejected invalid moves. Additional categories verify category independence. Unit tests cover ordering, atomicity, duplicate identities, custom tiers, escaping, and rendering.

Set EVAL_OUTPUT_DIR=/absolute/path to save the protocol test's preview sequence and board JSON. The sample rankings are test fixtures, not recommendations.

These deterministic tests validate the tool flow. They do not establish that every model researches good candidates, that a specific chat client's App sandbox permits downloads, or that a hosted deployment is working. Use the acceptance test below in your target chat client:

1. Give an unanticipated category; verify relevant options appear in chat without supplying the list yourself.

2. Rank two items; verify the image and card agree.

3. Correct one item; verify the other ranking is preserved.

4. Undo; verify the previous snapshot is restored.

5. Say download; verify an actual PNG attachment opens and matches the visible ranking.

Privacy and deployment

The application stores no board data. Logo lookup sends company domains to Google’s favicon service. MCP clients and hosting providers may retain tool-call logs; configure their retention separately. The unauthenticated endpoint exposes only pure rendering tools, but a public host should apply request limits and authentication appropriate to its audience. No TierMaker cookies or tokens belong in this project.

For browser interaction verification, with the inspector running on port 3417 and Python Playwright plus Chromium installed, run python3 tests/ui_smoke.py. This checks dropdown and drag ranking, preservation, undo, unranking, and the PNG preview/fallback in an isolated browser.

Optional local exports

Set TIER_LIST_EXPORT_DIR to a directory to save each requested PNG on the server as well as returning it through MCP. Filenames include a content hash so different boards do not overwrite each other. Exported files remain until you remove them; this is opt-in and is not a public download endpoint. The card’s Download in chat action includes its export filename and complete board so it does not depend on the host forwarding widget state.

Contributing

Issues and pull requests are welcome. Run npm run build, npm run typecheck, and npm test before submitting changes. Changes to the card should also pass python3 tests/ui_smoke.py against a running inspector on port 3417. CI exercises the eight-tool board with a host that omits null fields from responses, checking that unranked items survive every move.

License

MIT.