Iconmind MCP server

IconMind

Open-source icons for AI-era software. LLMs, agents, MCP, RAG, and everything around them.

Website · Browse all icons · Collections · Docs · MCP server

import { Agent, ContextWindow, VectorDatabase } from "@iconmind/react";

<Agent />
<ContextWindow variant="duotone" weight="bold" />
<VectorDatabase size={32} className="text-violet-500" />

5,287 icons · 31,722 drawings · 10 packages · one generated source of truth.

Every icon is a declaration compiled into six cells by a validator that refuses geometry it cannot draw correctly, and a nightly job rasterises all 31,722 of them and fails if any two icons render the same. That is how a set this size still reads as one hand, and why an assistant can be handed an MCP server that returns real names instead of .

Icons get redrawn when that scan or the contact sheet catches them — a whole round of rooms and appliances went back on the board because shapes were fusing at 16 px — and Bluetooth was dropped because its rune needs three stroke crossings where the set allows two. Everything else that was turned away is written down too.

Why this exists

Build a UI for an agent, a RAG pipeline, or an MCP server and you run out of icons almost immediately. There is no context-window, no reranker, no tool-calling, no mcp-resource in any mainstream set, so interfaces reach for a generic robot, a generic database, a generic lightning bolt, and the meaning is lost.

Lucide, Heroicons and Phosphor are excellent, and a thousand domain-specific AI icons would be scope creep for any of them. IconMind sits next to them, not instead of them. The 24px grid and 2px regular stroke match Lucide on purpose, so the two can share an interface without clashing.

The set

5,287 icons · 20 domains · 6 cells each. Two variants, outline and duotone, each drawn at three weights (1.5 / 2 / 2.5). A weight is a real drawing constraint, not a stroke-width slider: gap and legibility rules tighten as the stroke grows, and the validator checks every cell separately.

Duotone is derived, never drawn: every closed body tints at 20% behind the strokes, and icons made only of open marks (a check, an arrow) get a halo, the same paths echoed wider behind them, so the whole set answers the variant switch, not just the lucky icons.

Shortened. The full README is on GitHub.