Coding agents

Agents often guess icon names wrong. Give Claude Code, Cursor or Codex the real set through MCP, a CLI or a skill.

MCP server

@coolicons/mcp lets your agent search the set, read an icon's SVG and get the right import before it writes code. It needs no API key and no network: the whole set ships in the package.

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

Run it in your project folder. Add --scope user before coolicons to use it in every project.

MCP tools

Tool	What it does
`search_icons`	Finds icons by name or meaning, like delete for trash. Several searches in one call, separated by commas.
`get_icon`	SVG source of one icon in any style and corner radius, with size and stroke width.
`get_usage`	Install command and import for React, Vue, SVG or shadcn/ui, with the right entry for the style.
`describe_set`	Styles, corners, categories, packages and naming rules.

A name that does not exist comes back with the closest names, so a typo resolves in one step. Names that need the Icon suffix get it in every import.

Command line

@coolicons/cli does the same in a terminal, for agents without MCP and for your own scripts.

Terminal

npx @coolicons/cli search delete
npx @coolicons/cli show trash --style fill > trash.svg
npx @coolicons/cli add trash bell --style bulk --radius round --out src/icons
npx @coolicons/cli list --category Arrow

Add --json to search and list for output a script can read. --help lists every option.

Skill

The coolicons skill teaches your agent the naming rules, the styles and the imports, and tells it to look up every name instead of guessing. Install it into any agent that reads skills:

Terminal

npx skills add krystonschwarze/coolicons

The skill uses the MCP server or the command line when they are there and the full icon list when they are not.

llms.txt

For an agent that reads from the web, point it to coolicons.cool/llms.txt. It covers install, styles and names in one page. Every icon with its tags is in llms-full.txt.

Which one

Use the MCP server if your agent supports it. Use the command line if it only runs shell commands. Add the skill to either one, so the agent knows when to look up a name. llms.txt is for agents that can only read a URL.