Autonomous Coding Agent

A long-running autonomous coding agent powered by the Claude Agent SDK. Builds complete applications over multiple sessions using a three-agent pattern with support for multiple technology stacks.

Based on leonvanzyl's autonomous-coding

Features

Multi-session development - Maintains progress across context windows via SQLite database and git
Three agent types - Initializer, Enhancement, and Coding agents for different phases
Multiple tech stacks - Node.js/React (default) or Laravel with React/Vue/Blade
Post-build enhancements - Add new features to existing projects without starting over
Browser automation testing - Playwright MCP for end-to-end verification
Security sandboxing - Restricted filesystem and command allowlist

Prerequisites

Claude Code CLI (Required)

macOS / Linux:

curl -fsSL https://claude.ai/install.sh | bash

Windows (PowerShell):

irm https://claude.ai/install.ps1 | iex

Authentication

Claude Pro/Max Subscription - Use claude login to authenticate (recommended)
Anthropic API Key - Pay-per-use from https://console.anthropic.com/

For Laravel Projects (Optional)

PHP 8.2+
Composer
Node.js 18+

Quick Start

1. Clone the Repository

git clone https://github.com/jacob8919/autonomous-coding.git
cd autonomous-coding

2. Run the Start Script

Windows:

start.bat

macOS / Linux:

./start.sh

The start script will:

Check if Claude CLI is installed
Check authentication (prompt for claude login if needed)
Create Python virtual environment
Install dependencies
Launch the main menu

3. Main Menu Options

[1] Create new project          - Start fresh with AI-assisted spec generation
[2] Continue existing project   - Resume work on a previous project
[3] Add features to existing    - Add new features after initial build
[q] Quit

Technology Stacks

When creating a new project, you can choose from:

Stack	Frontend	Backend	Database
Node.js + React (default)	React	Express	SQLite
Laravel + React	React/Inertia	Laravel	SQLite
Laravel + Vue	Vue/Inertia	Laravel	SQLite
Laravel + Blade	Blade/Alpine.js	Laravel	SQLite

Laravel stacks include automatic setup via Composer and Laravel Breeze, plus Laravel Boost MCP integration.

How It Works

Three-Agent Pattern

┌─────────────────────────────────────────────────────────────────┐
│                    AUTONOMOUS CODING WORKFLOW                    │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  [1] INITIALIZER AGENT (First Session)                          │
│      • Reads app_spec.txt                                        │
│      • Creates features in SQLite database                       │
│      • Sets up project structure                                 │
│      • Initializes git repository                                │
│                          ↓                                       │
│  [2] CODING AGENT (Subsequent Sessions)                          │
│      • Gets next feature from database                           │
│      • Implements feature                                        │
│      • Tests with browser automation                             │
│      • Marks feature as passing                                  │
│      • Commits progress                                          │
│                          ↓                                       │
│  [3] ENHANCEMENT AGENT (When adding features)                    │
│      • Reads enhancement_spec.txt                                │
│      • Checks for duplicate features                             │
│      • Adds new features to database                             │
│      • Returns to Coding Agent                                   │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Session Management

Each session runs with a fresh context window
Progress persisted via features.db (SQLite) and git commits
Agent auto-continues between sessions (3 second delay)
Press Ctrl+C to pause; run start script again to resume

Adding Features After Build

Already built a project but need more features? Use menu option [3]:

Select "Add features to existing project"
Choose your project
Claude Code launches with /add-features command
Describe the new features you want
Choose priority (before or after pending features)
Exit and select "Continue existing project"
Enhancement Agent adds features, then Coding Agent implements them

MCP Servers

The agent uses Model Context Protocol (MCP) servers for extended capabilities:

Playwright (Browser Automation)

Navigate, click, type, screenshot
Form filling and element interaction
Console and network monitoring
End-to-end feature verification

Features (Database Management)

feature_get_stats - Progress statistics
feature_get_next - Next feature to implement
feature_mark_passing - Mark feature complete
feature_create_bulk - Add multiple features
feature_search - Find existing features

Laravel Boost (Laravel Projects Only)

database-schema - Inspect database structure
list-routes - View application routes
tinker - Run Tinker commands
search-docs - Query Laravel documentation
last-error - Get recent errors
And 11 more tools...

Project Structure

autonomous-coding/
├── start.bat / start.sh          # Platform start scripts
├── start.py                      # Main menu and project management
├── autonomous_agent_demo.py      # Agent entry point
├── agent.py                      # Session orchestration
├── client.py                     # Claude SDK configuration
├── security.py                   # Command allowlist
├── prompts.py                    # Prompt loading
├── progress.py                   # Progress tracking
├── api/
│   ├── database.py               # SQLAlchemy models
│   └── migration.py              # Schema migrations
├── mcp_server/
│   └── feature_mcp.py            # Feature management MCP
├── stacks/                       # Laravel stack configs
│   ├── laravel-react.md
│   ├── laravel-vue.md
│   └── laravel-blade.md
├── .claude/
│   ├── commands/
│   │   ├── create-spec.md        # /create-spec command
│   │   └── add-features.md       # /add-features command
│   └── templates/                # Agent prompt templates
└── generations/                  # Generated projects

Generated Project Structure

generations/my_project/
├── features.db                   # SQLite database (source of truth)
├── prompts/
│   ├── app_spec.txt              # Application specification
│   ├── initializer_prompt.md     # First session prompt
│   ├── coding_prompt.md          # Coding session prompt
│   └── enhancement_spec.txt      # (If adding features)
├── .claude/
│   └── mcp_servers.json          # Project MCP config (Laravel)
├── init.sh                       # Environment setup
├── claude-progress.txt           # Session notes
└── [application files]           # Generated code

Timing Expectations

Phase	Duration
Initialization	10-20+ minutes (generating features)
Each coding session	5-15 minutes per feature
Full application	Hours across multiple sessions

The feature count determines scope. Typical ranges:

Simple apps: 20-50 features
Medium apps: 100-150 features
Complex apps: 200+ features

Security Model

Defense-in-depth approach (see security.py):

OS-level Sandbox - Bash commands run isolated
Filesystem Restrictions - Operations restricted to project directory
Command Allowlist - Only permitted commands execute:

Category	Commands
File inspection	`ls`, `cat`, `head`, `tail`, `wc`, `grep`
Node.js	`npm`, `npx`, `node`
PHP/Laravel	`php`, `composer`
Version control	`git`
Process mgmt	`ps`, `lsof`, `sleep`, `pkill`

Configuration

N8N Webhook (Optional)

Send progress notifications to N8N:

# .env file
PROGRESS_N8N_WEBHOOK_URL=https://your-n8n.com/webhook/id

Payload:

{
  "event": "test_progress",
  "passing": 45,
  "total": 200,
  "percentage": 22.5,
  "project": "my_project"
}

Customization

Application Specification

Use /create-spec for interactive creation
Or edit prompts/app_spec.txt directly

Adding Commands

Edit ALLOWED_COMMANDS in security.py

Custom MCP Servers

Add to .claude/mcp_servers.json in project directory

Troubleshooting

Issue	Solution
"Claude CLI not found"	Install CLI per Prerequisites section
"Not authenticated"	Run `claude login`
"Appears to hang on first run"	Normal - generating features takes time. Watch for `[Tool: ...]` output
"Command blocked"	Add to `ALLOWED_COMMANDS` in `security.py`
"Permission denied for MCP tool"	Add tool to `client.py` allowed list

License

Internal Anthropic use.

Name		Name	Last commit message	Last commit date
Latest commit History 35 Commits
.claude		.claude
api		api
mcp_server		mcp_server
stacks		stacks
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
agent.py		agent.py
autonomous_agent_demo.py		autonomous_agent_demo.py
client.py		client.py
progress.py		progress.py
prompts.py		prompts.py
requirements.txt		requirements.txt
security.py		security.py
start.bat		start.bat
start.py		start.py
start.sh		start.sh
test_security.py		test_security.py

jacob8919/autonomous-coding

Folders and files

Latest commit

History

Repository files navigation