Agent Builder

A web UI for creating and managing Claude Code subagents.

Features

• Create agents with a visual form - name, description, tools, model, color
• Generate prompts using claude --print opus for intelligent prompt creation
• Preview agent files before saving
• Configure hooks for PreToolUse, PostToolUse, and Stop events
• Edit and delete existing agents
• Dark theme GitHub-inspired UI

Quick Start

Local Development

# Option 1: Use the start script (handles venv)
./start.sh

# Option 2: Manual setup
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python server.py

Open http://localhost:8086

Docker

# Build and run
docker-compose up -d

API Endpoints

Method	Endpoint	Description
GET	/api/agents	List all agents
GET	/api/agents/{name}	Get agent details
POST	/api/agents	Create new agent
PUT	/api/agents/{name}	Update agent
DELETE	/api/agents/{name}	Delete agent
POST	/api/generate-prompt	Generate prompt with Claude
POST	/api/preview	Preview agent file
GET	/api/health	Health check

Agent File Format

Agents are stored as Markdown files with YAML frontmatter in ~/.claude/agents/:

---
name: my-agent
description: What the agent does
tools: Bash, Read, Write
model: haiku
color: blue
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate.sh"
---

System prompt content here...

Available Tools

• Bash - Execute shell commands
• Read - Read files
• Write - Create/overwrite files
• Edit - Edit existing files
• Glob - Find files by pattern
• Grep - Search file contents
• Task - Spawn subagents
• WebSearch - Search the web
• WebFetch - Fetch web pages
• NotebookEdit - Edit Jupyter notebooks
• AskUserQuestion - Ask user questions
• EnterPlanMode - Enter planning mode
• ExitPlanMode - Exit planning mode

Available Models

• haiku - Fast, low cost
• sonnet - Balanced
• opus - Most capable
• inherit - Use parent model

Hook Events

• PreToolUse - Before tool execution (can block)
• PostToolUse - After tool execution
• Stop - When agent finishes

Best Practices

1. Write clear descriptions - Claude uses them to decide when to delegate
2. Limit tool access - Only grant necessary tools
3. Use haiku for simple tasks - Save cost and reduce latency
4. Add hooks for validation - Catch errors early
5. Include "use proactively" in descriptions for auto-delegation

License

MIT