You can add custom tools to Claude Code through two mechanisms: the plugin system (for tools bundled directly into Claude Code’s runtime) and MCP servers (for tools exposed by external processes).
The easiest way to give Claude new capabilities is through an MCP server. MCP tools appear in Claude’s tool list just like built-in tools.
Advantages:
- No changes to Claude Code itself
- Works with any language (Node.js, Python, Go, Rust, etc.)
- Hot-reloadable without restarting Claude Code
- Shareable across teams via a central MCP server
See MCP Protocol for a complete guide to building MCP servers.
Plugins can register tools that run inside the Claude Code process. This is useful for tools that need deep integration with Claude Code’s internals.
Create a plugin in ~/.claude/plugins/my-plugin/:
my-plugin/
├── manifest.json
├── index.js
└── tools/
└── myCustomTool.js
{
"name": "my-plugin",
"version": "1.0.0",
"tools": ["tools/myCustomTool.js"]
}
module.exports = {
name: "my_custom_tool",
description: "Does something custom",
inputSchema: {
type: "object",
properties: {
input: { type: "string", description: "The input value" }
},
required: ["input"]
},
async call({ input }) {
// Tool implementation
return { output: `Processed: ${input}` };
}
};
- Write clear
description fields — Claude uses them to decide when to use the tool
- Use
required to mark mandatory fields
- Use
enum to constrain allowed values
- Add
default values where appropriate
Claude Code also has an AskUserQuestionTool for tools that need to interactively ask the user for input mid-execution. Custom tools can trigger this pattern by returning a question response that Claude translates into a user prompt.
For most use cases, MCP is the better choice for custom tools. The plugin API is more powerful but more complex, and it requires familiarity with Claude Code’s internals.