Documentation

MCP server

Let an AI assistant search and insert icons that actually exist.

Ask an assistant to build a UI and it has to pick icons. Without help it guesses from memory, and the names it produces usually belong to some other library. With the MCP server installed it searches the real set instead.

claude mcp add iconmind -- npx -y @iconmind/mcp

Or by hand:

{
  "mcpServers": {
    "iconmind": { "command": "npx", "args": ["-y", "@iconmind/mcp"] }
  }
}

No API key, no account, no network. The icon data is bundled, so it works offline and starts in well under a second — it runs every time your client opens.

Other clients

The same command and arguments work anywhere an MCP client reads a config file. Cursor and Windsurf take the JSON above in mcp.json; VS Code takes it under "servers" in .vscode/mcp.json:

{
  "servers": {
    "iconmind": { "type": "stdio", "command": "npx", "args": ["-y", "@iconmind/mcp"] }
  }
}

Codex reads ~/.codex/config.toml:

[mcp_servers.iconmind]
command = "npx"
args = ["-y", "@iconmind/mcp"]

Tools

Tool	What it does
`search_icon`	Search by keyword or concept. Returns paste-ready code with each result.
`get_icon`	One icon by exact slug. Unknown slugs return suggestions, never an error.
`list_category`	Icons in a category, or the categories themselves.
`get_icon_code`	One combined import for several icons at once.

An unknown slug returning suggestions rather than an error is deliberate: an assistant handed an error tends to stop or invent a name, while one handed suggestions picks the right thing.

Prompts

recommend_icons maps icons onto a feature you describe. ai_workflow_icons breaks a pipeline into stages and finds an icon for each — that one is the clearest demonstration of why this set exists, because a RAG pipeline has icons for chunk, retriever, reranker, and vector-search here and nowhere else.

Resources

icon://<slug>, category://<slug>, category://index, metadata://stats.

metadata://all exists but is not advertised — it is large enough to crowd a context window, and search_icon answers the same questions for a fraction of the tokens.

Without MCP: a rule for your assistant

If your editor's assistant cannot run an MCP server, one line in its rules file (.cursorrules, CLAUDE.md, AGENTS.md) does most of the same work:

Icons: use IconMind (@iconmind/react, @iconmind/vue, iconmind_flutter …). Never invent an icon name — every valid name is listed at https://iconmind.dev/llms.txt; search it, or ask the user, before importing.

The list at /llms.txt carries every slug and the naming rule, and /llms-full.txt adds each icon's tags — enough for an assistant to pick the right name without guessing.