DDEV AI Workspace – AI-Assisted Drupal Development

One command installs a complete AI development stack inside your existing DDEV project: OpenCode and Claude Code in their own containers, a headless browser, a task tracker, an autonomous task runner, and a pack of 11 Drupal agents, 15 rules and 38 skills that all of them share. A separate host package, ddev-ai-arena, adds parallel candidates in isolated copies of the project.

Built for Drupal. The agents understand Drupal 10/11 APIs, coding standards, caching, render arrays, testing and the module/theme ecosystem out of the box.

Created by Robert Menetray · Sponsored by DruScan — Drupal site auditing and monitoring tools.

At a glance

Everything runs in containers on the DDEV network. AI containers never touch the database directly: they reach the web container over SSH (ddev-ai-ssh) to run drush, composer, PHPUnit and PHPStan.

Prerequisites

• DDEV >= v1.24.10
• An Anthropic subscription or API key for Claude Code, or any provider OpenCode supports (Anthropic, OpenAI, OpenCode Zen, free models, a LiteLLM proxy)

Quick Start

cd /path/to/your-drupal-project
ddev add-on get trebormc/ddev-ai-workspace
ddev restart

ddev claude-code   # or: ddev cc
ddev opencode      # or: ddev oc

Each tool guides you through authentication the first time. Credentials live in shared directories on your host, so every DDEV project reuses them:

Then, when you need them:

ddev ralph --backend claude          # autonomous loop over requirements.md and Beads tasks
ddev bd ready                        # task tracker

Architecture

HOST
 ├─ your AI harness (Claude Code / OpenCode) in a coordinator project ── ai-arena ──▶ candidate copies of this project
 └─ ai-notify-bridge (optional desktop notifications, port 5454)

DDEV Network (one project)
 ├─ claude-code ─┬─ SSH ──▶ web (PHP, Drupal, drush, composer, phpunit, phpstan)
 ├─ opencode ────┤─ SSH ──▶ beads (bd task tracker, .beads/ in the project)
 │               └─ HTTP MCP ─▶ playwright-mcp (headless Chromium)
 ├─ ralph ── SSH ─▶ claude-code or opencode (delegates every AI step)
 └─ agents-sync ─▶ named volumes read by claude-code and opencode

The three ways to work

Interactive. Open ddev claude-code or ddev opencode and talk to the orchestrator. It runs on a cheap model and delegates to specialists (backend, theme, tests, a seven-judge code review, a context-firewall explorer) that run on the right model for the job. Task playbooks (bug fix, feature, refactoring, investigation) and one verification bar apply to every task: tests alone are not enough, behavior is observed on the real site and the evidence is in the summary.

Autonomous. Write requirements.md, run ddev ralph. Ralph plans Beads tasks, then executes them one by one through an AI container until none is left, leaving a decision trail in the task notes. Audit fixers and test generators ship ready to run.

Arena. From the AI harness on your host, opened in a small coordinator project that sits above your environments (a generic DDEV project with this workspace installed, no Drupal), ask for an arena. ai-arena creates one git worktree with its own DDEV project and database per candidate (each a copy of this workspace), runs each candidate detached in its own AI container (different harnesses and models if you like), verifies them mechanically (site bootstrap, PHPStan, PHPCS, PHPUnit), lets a project-manager session judge the diffs blind, and promotes the winner to a feature/ branch. Questions bubble up to you only when a human must decide. It is a host package installed once with its own install.sh, not a DDEV add-on.

Why both OpenCode and Claude Code?

Claude Code works with an Anthropic subscription (Max plan) and gives access to the most capable models (Opus, Sonnet). The subscription cannot be used from other tools. Best for architecture decisions, large refactors and multi-file changes.

OpenCode is open source and connects to many providers, including free models and OpenCode Zen. Great to keep costs low, to run local models, and as a second family of models for arena candidates and cross-tool second opinions.

Having both in the same project means you pick per task, keep working if one provider is down, and avoid vendor lock-in. Install only one if you prefer (see below).

Individual installation

ddev add-on get trebormc/ddev-opencode      # OpenCode (+ ai-ssh, beads, playwright-mcp, agents-sync)
ddev add-on get trebormc/ddev-claude-code   # Claude Code (same dependencies)
ddev add-on get trebormc/ddev-ralph         # Ralph (needs OpenCode or Claude Code first)

Desktop notifications (optional, Linux only)

Containers can notify you when a session finishes or needs attention. The bridge is a host-level tool installed once for all projects:

curl -fsSL https://raw.githubusercontent.com/trebormc/ai-notify-bridge/main/install.sh | bash

It runs a small HTTP server on port 5454 as a systemd user service. Without it, containers get an instant “connection refused” and nothing else changes. See ai-notify-bridge.

Repositories

This workspace is 10 independent git repositories. Install them all with this meta add-on or pick the ones you need.

AI tools

Shared dependencies

Agents

Uninstallation

ddev add-on remove ddev-ai-workspace
ddev restart

The meta add-on only groups the others: remove them individually with ddev add-on remove <name> if you want them gone too.

Disclaimer

This project is an independent initiative by Robert Menetray, sponsored by DruScan. It is not affiliated with, endorsed by, or sponsored by Anthropic (Claude Code), OpenCode, Beads, Playwright, Microsoft, or DDEV. These are third-party tools integrated here for convenience.

AI-generated code may contain errors, security issues, or unintended behavior. Always review AI-generated changes before deploying to production. Unattended runs (Ralph, arena candidates) should be followed by a thorough human review. The author assumes no responsibility for damages caused by the use or misuse of these configurations or the code they produce.

License

Apache-2.0 (all repositories)