Brain Icon Visual MCP Server

mcp-icon-visual

MCP server that gives AI agents the ability to search, retrieve, compare, and render SVG icons. Built with Rust and the rmcp SDK.

Combines local filesystem access with a remote rendering API to resolve SVGs from any source — inline strings, URLs, local files, or Iconify IDs — and render them onto visual comparison grids.

The key insight: every tool returns images. A vision-capable LLM doesn't just get metadata — it sees the icons. This turns icon selection from guesswork into a visual conversation.

search_icons("machine learning") — 4x4 grid, cells 0–15, returned as PNG

Agent Pipelines

These tools are designed to be composed by AI agents with vision. Below are three proven patterns.

1. Visual Search & Selection

An agent searches for icons and visually picks the best match — no blind URL selection.

                          ┌─────────────────────────────┐
                          │        Agent + Vision        │
                          └──────────────┬──────────────┘
                                         │
                        ┌────────────────┼────────────────┐
                        v                v                v
                  search_icons     search_icons     search_icons
                  "cloud"          "upload"          "storage"
                        │                │                │
                        v                v                v
                  ┌──────────┐    ┌──────────┐    ┌──────────┐
                  │ 4x4 grid │    │ 4x4 grid │    │ 4x4 grid │
                  │  (sees)  │    │  (sees)  │    │  (sees)  │
                  └────┬─────┘    └────┬─────┘    └────┬─────┘
                       │               │               │
                       └───────────────┼───────────────┘
                                       v
                              "cells 3, 7, 12 match
                               the project style"
                                       │
                                       v
                               render_grid [3 winners]
                                       │
                                       v
                              ┌──────────────────┐
                              │  final comparison │
                              │   grid (sees)     │──> pick best
                              └──────────────────┘

The agent runs multiple searches, looks at each grid, picks candidates by cell number, then renders a final side-by-side comparison to make the choice. Works especially well when the agent has context about the project's visual style.

2. Iterative Icon Editing

An agent modifies SVG code and uses render_grid as a visual diff to track progress across iterations.

         get_svg "mdi/cloud"
                │
                v
          ┌───────────┐
          │ raw SVG    │
          │ markup     │
          └─────┬─────┘
                │
       ┌────────┴─────────────────────────────────────┐
       │              Edit Loop                        │
       │                                               │
       │   Agent modifies SVG code                     │
       │        │                                      │
       │        v                                      │
       │   render_grid [original, v1, v2, v3]          │
       │        │                                      │
       │        v                                      │
       │   ┌──────────────────────────────┐            │
       │   │ cell 0: original             │            │
       │   │ cell 1: thicker strokes      │            │
       │   │ cell 2: rounded corners      │            │
       │   │ cell 3: filled variant       │            │
       │   └──────────────┬───────────────┘            │
       │                  │                            │
       │        agent sees all versions                │
       │        decides next edit                      │
       │                  │                            │
       │                  └──── loop ──────────────────┘
       │
       v
  save final SVG to disk

Each iteration, the agent passes all versions (original + edits) as inline SVGs to render_grid. It sees them side-by-side in one image and decides whether to keep iterating or stop.