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. The original always stays in cell 0 as a reference.

3. Style Matching & Consistency

An agent ensures new icons match the visual style of existing project icons.

  Project icons on disk             Candidate sources
  ─────────────────────             ─────────────────
  /app/icons/nav-home.svg           search_icons "settings"
  /app/icons/nav-search.svg              │
  /app/icons/nav-profile.svg             v
          │                        picks cells 2, 5, 9
          │                              │
          └──────────┬───────────────────┘
                     v
              render_grid [
                /app/icons/nav-home.svg,      ← cell 0: existing
                /app/icons/nav-search.svg,    ← cell 1: existing
                /app/icons/nav-profile.svg,   ← cell 2: existing
                candidate_url_1,              ← cell 3: candidate
                candidate_url_2,              ← cell 4: candidate
                candidate_url_3,              ← cell 5: candidate
                "<svg>...custom edit...</svg>" ← cell 6: agent's edit
              ]
                     │
                     v
              ┌────────────────────────────────┐
              │ grid image:                     │
              │  existing icons  vs  candidates │
              │  (agent sees style match/clash) │
              └────────────────┬───────────────┘
                               │
                      agent judges:
                  "cell 4 matches stroke weight
                   and corner radius of cells 0-2,
                   cell 3 is too thin,
                   cell 5 wrong fill style"
                               │
                               v
                       get_svg cell_4_url
                               │
                               v
                    save to /app/icons/nav-settings.svg

By mixing local file paths and remote sources in a single grid, the agent can visually compare existing project icons against candidates in one shot. The grid becomes a style audit tool.

Combining Pipelines

These patterns compose naturally. A real-world workflow might:

1. Search across multiple keywords (pipeline 1)
2. Compare top candidates against existing project icons (pipeline 3)
3. Edit the closest match to fix style inconsistencies (pipeline 2)
4. Verify the final icon one more time against the full icon set (pipeline 3)

All driven by a single agent with vision, using three tools.

Tools

search_icons

Semantic icon search. Returns a 4x4 PNG grid (512x512, cells 0-15) and a mapping of cell numbers to SVG URLs.

Parameter	Type	Default	Description
keyword	string	required	Semantic search query (e.g. "cloud computing", "arrows")
background	string	"#FFFFFF"	Grid background color, CSS hex
color	string | null	null	Recolor all icons to this CSS hex color

get_svg

Retrieve raw SVG markup from any source. Returns the complete SVG string for inspection or modification.

Parameter	Type	Description
source	string	URL, Iconify ID (mdi/cloud), absolute file path, or inline <svg>

render_grid

Render 1-16 SVGs onto a 4x4 comparison grid (512x512, cells 0-15). Useful for comparing icon variants side-by-side or previewing local icons alongside search results.

Parameter	Type	Default	Description
sources	string[]	required	1-16 SVG sources (URLs, Iconify IDs, file paths, or inline SVGs)
background	string	"#FFFFFF"	Grid background color, CSS hex

Source Resolution

All tools share a unified resolution pipeline that classifies sources automatically:

Source type	Example	Resolution
Inline SVG	<svg xmlns="...">...</svg>	Returned as-is
URL	https://mdn.alipayobjects.com/.../original	Resolved via API
File path	/home/user/icons/logo.svg	Read from local disk
Iconify ID	mdi/cloud, lucide/home	Resolved via API

File paths are detected by prefix (/, ./, ~/) or .svg extension. Everything else without http:///https:// is treated as an Iconify ID.

Grid Format

All grid outputs are 512x512 PNG images with a 4x4 layout (128x128 cells), numbered 0-15 left-to-right, top-to-bottom:

┌────┬────┬────┬────┐
│  0 │  1 │  2 │  3 │
├────┼────┼────┼────┤
│  4 │  5 │  6 │  7 │
├────┼────┼────┼────┤
│  8 │  9 │ 10 │ 11 │
├────┼────┼────┼────┤
│ 12 │ 13 │ 14 │ 15 │
└────┴────┴────┴────┘

Cell numbers appear at the bottom-left of each cell. Unused cells show the number only.

Installation

npx (recommended)

No build needed. Works on Linux (x64/arm64), macOS (Intel/Apple Silicon), and Windows (x64).

{
  "mcpServers": {
    "icon-visual": {
      "command": "npx",
      "args": ["-y", "@br-ai-n/mcp-icon-visual"]
    }
  }
}

From source

Requires Rust (edition 2024):

cargo build --release --target x86_64-unknown-linux-musl

The release profile produces a fully static binary (musl libc, rustls-tls, LTO, stripped).

Usage

# Default upstream API
./mcp-icon-visual

# Custom API endpoint
./mcp-icon-visual --api-base http://localhost:3000

The server communicates over stdio using the MCP protocol.

MCP Client Configuration

Claude Desktop / Cursor

{
  "mcpServers": {
    "icon-visual": {
      "command": "npx",
      "args": ["-y", "@br-ai-n/mcp-icon-visual"]
    }
  }
}

Custom API endpoint

{
  "mcpServers": {
    "icon-visual": {
      "command": "/path/to/mcp-icon-visual",
      "args": ["--api-base", "http://localhost:3000"]
    }
  }
}

Project Structure

src/
├── main.rs           # Entrypoint, CLI args, stdio transport
├── server.rs         # IconServer + MCP tool routing + response builders
├── api_client.rs     # HTTP client for upstream rendering API
├── resolve.rs        # Source classification and SVG resolution
├── error.rs          # Error types
└── tools/
    ├── mod.rs
    ├── search_icons.rs   # SearchIconsParams
    ├── get_svg.rs        # GetSvgParams
    └── render_grid.rs    # RenderGridParams

Testing

# Unit tests (source classification)
cargo test

# Integration tests (requires network access to icons.buan.me)
cargo test -- --ignored

Dependencies

Crate	Purpose
rmcp	MCP server SDK (tool routing, stdio transport)
tokio	Async runtime
reqwest	HTTP client (rustls-tls, no OpenSSL)
serde / serde_json	Serialization
schemars	JSON Schema generation for tool parameters
clap	CLI argument parsing
base64	Base64 encoding
tracing	Logging
thiserror	Error derive macros

License

MIT License

Copyright (c) 2025

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.