Summary

AutoCoder is an autonomous coding agent using the Claude SDK to build complete applications over multiple sessions. It employs a two-agent pattern and offers a real-time React UI for progress monitoring, targeting developers seeking automated application development workflows.

How It Works

A two-agent system orchestrates development: an Initializer Agent sets up project structure, features in SQLite, and Git. Subsequent Coding Agent sessions iteratively implement features, persisting progress via SQLite and Git. Features are managed through an MCP server with tools for status tracking and updates, enabling long-running, stateful development cycles.

Quick Start & Requirements

• Prerequisites: Claude Code CLI (install via curl or PowerShell script).
• Authentication: Claude Pro/Max (claude login) or Anthropic API Key.
• Installation: Handled by start.bat / start.sh scripts, which set up a Python venv and install dependencies.
• Running: Web UI (start_ui.bat / ./start_ui.sh at http://localhost:5173) or CLI (start.bat / ./start.sh).
• Project Creation: Interactive specification generation via /create-spec.
• Links: Video Tutorial (implied), Claude Code CLI installation scripts.

Highlighted Details

• Real-time UI: React frontend with project selection, Kanban-style feature tracking, and live agent output streaming.
• Security Model: Defense-in-depth with OS sandboxing, filesystem restrictions, and a strict bash command allowlist (security.py).
• Tech Stack: React 18 (TS), TanStack Query, Tailwind CSS v4, Radix UI, FastAPI, SQLite.
• Generated Projects: Includes features.db for state, prompt files, and an initialization script (init.sh).

Maintenance & Community

Copyrighted by Leon van Zyl. No specific community channels or roadmap links are detailed in the README.

Licensing & Compatibility

Licensed under GNU Affero General Public License v3.0 (AGPL-3.0). This strong copyleft license requires derivative works or network-distributed modifications to be released under the same license, impacting commercial or closed-source integration.

Limitations & Caveats

Application building is time-intensive, requiring many hours across multiple sessions. The initial agent session may appear unresponsive for minutes during test case generation. The security model's command allowlist may block intended agent actions if not explicitly permitted.