claude-code

Claude Code Agent

An Agentuity agent that demonstrates Claude Agent SDK integration with Agentuity Sandboxes for conversational code intelligence and safe code execution.

What It Does

• Code Intelligence — Uses Claude Agent SDK (@anthropic-ai/claude-agent-sdk) to read, analyze, and generate code in a local workspace seeded with sample TypeScript files.
• Sandbox Execution — Runs user code safely in isolated Agentuity Sandboxes (ctx.sandbox.run()) using the Bun runtime.
• Conversation History — Maintains multi-turn conversations via Agentuity thread state, so the agent remembers prior context.
• Chat Frontend — A React-based chat UI for interacting with the agent conversationally.

Quick Start

1. Import the project: agentuity project import
2. Add ANTHROPIC_API_KEY=... to .env
3. Install dependencies: bun install
4. Start the dev server: bun run dev
5. Open http://localhost:3500 and start chatting about code.

Project Structure

claude-code/
├── src/
│   ├── agent/claude-code/
│   │   ├── index.ts         # Agent handler (Claude Agent SDK + Sandbox)
│   │   └── sample-files.ts  # Sample TypeScript files for the workspace
│   ├── api/
│   │   └── index.ts         # Chat API routes
│   └── web/
│       ├── App.tsx           # Chat UI
│       ├── App.css           # Tailwind theme
│       ├── frontend.tsx      # React entry point
│       └── index.html        # HTML template
├── app.ts                    # Application entry point
├── vite.config.ts            # Vite config (React + Tailwind)
└── package.json

How It Works

1. The frontend sends a prompt via POST /api/chat.
2. The agent initializes a workspace with sample files and calls Claude Agent SDK's query().
3. Claude Code reads, writes, and analyzes files using built-in tools (Read, Write, Edit, Glob, Grep).
4. When the user asks to run code, the agent ships workspace files to an Agentuity Sandbox for execution.
5. Results (analysis + execution output) are returned to the chat UI.

Security Considerations

The agent calls Claude Agent SDK with permissionMode: 'bypassPermissions' and allowDangerouslySkipPermissions: true. This is intentional for this example:

• Tool scope is narrow. Claude Code is granted only Read, Write, Edit, Glob, and Grep. There is no Bash tool, so it cannot execute arbitrary commands.
• Workspace is isolated. Each thread gets its own temporary directory; there is no shared filesystem between sessions.
• Code execution is sandboxed. When the user asks to run code, files are shipped to an Agentuity Sandbox (ctx.sandbox.run()), not executed on the host.

In a multi-tenant or shared environment, consider restricting cwd to a tighter boundary and auditing which tools are permitted before adopting this pattern.

API Endpoints

Method	Path	Description
POST	/api/chat	Send a chat message
GET	/api/chat/history	Get conversation history
DELETE	/api/chat/history	Clear conversation history

Sample Files

The agent workspace includes these reference files:

• fibonacci.ts — Recursive, iterative, and generator Fibonacci implementations
• math-tricks.ts — Prime checking, factorial, GCD, LCM utilities
• class-example.ts — Animal/Dog/Cat class hierarchy with inheritance
• hello.ts — Greeting and Fibonacci example

Requirements

• Bun v1.0+
• Agentuity CLI
• ANTHROPIC_API_KEY in local .env for the Claude Agent SDK

Environment

This example calls @anthropic-ai/claude-agent-sdk directly, so ANTHROPIC_API_KEY is the main manual local setup step.

Agentuity services such as dev mode, thread state, and sandboxes use AGENTUITY_SDK_KEY through the normal project setup from agentuity project import.

Minimal local .env:

ANTHROPIC_API_KEY=...

Resources

• Claude Agent SDK overview