claude_subagents

Configuration and agents for Claude Code, synced across machines via this repo.

Quick setup on a new machine

git clone https://github.com/hydrobiont/claude_subagents.git ~/claude_subagents
cd ~/claude_subagents
./install.sh

Then follow the manual steps printed at the end.

---

What's in this repo

agents/          Custom sub-agents (linked into ~/.claude/agents)
                   browser-operator    Browser automation via a dedicated, persistent Chromium profile
skills/          Custom skills (linked into ~/.claude/skills)
                   technical-writer/   dba-docs-internal + trainings2025 writer
scripts/
  browser-profile-setup.sh  Bootstrap helper — opens the persistent Chromium profile
                            so the user can log into a site once and reuse cookies forever
config/
  claude-settings.json   ~/.claude/settings.json — plugins + settings
  mcp.json               ~/.claude/.mcp.json template — MCP server definitions
memory/
  de-website/
    MEMORY.md    Project memory for dataegret.com website work
install.sh       Idempotent setup script

---

What install.sh does

Step	What	Where
Agents symlink	~/.claude/agents → ~/claude_subagents/agents	automatic
Skills symlink	~/.claude/skills → ~/claude_subagents/skills	automatic
Claude settings	copies config/claude-settings.json → ~/.claude/settings.json	automatic (skips if exists)
MCP config	merges config/mcp.json into ~/.claude/.mcp.json (adds new servers, never overwrites existing)	automatic
Browser profile dir	creates ~/.claude/chrome-profile/playwright/ (empty)	automatic
Project memory	copies memory/de-website/MEMORY.md into ~/.claude/projects/	automatic (skips if exists)
analytics-mcp	pipx install analytics-mcp	automatic if pipx available
Plugins	prints claude plugin install commands	manual (require interactive approval)

---

MCP servers

memory

Global conversation memory across sessions.

• Package: @modelcontextprotocol/server-memory (npx)
• No credentials needed.

analytics

Google Analytics 4 read access (run reports, get property details, etc.).

• Package: analytics-mcp (Python, installed via pipx)
• Auth: Google Application Default Credentials

# Install gcloud if needed
brew install --cask google-cloud-sdk

# Authenticate (opens browser)
gcloud auth application-default login

dataegret-staging

WordPress REST API access to dev2.dataegret.com.

• Package: @instawp/mcp-wp (npx)
• Credentials: fill in ~/.claude/.mcp.json after running install.sh

WORDPRESS_USERNAME  — your staging WP username
WORDPRESS_PASSWORD  — WordPress Application Password
                      (WP Admin → Users → Edit user → Application Passwords)

playwright-persistent

Playwright MCP server backed by a dedicated, persistent Chromium profile at ~/.claude/chrome-profile/playwright/. Used exclusively by the browser-operator sub-agent. Survives restarts, isolated from your daily Chrome, never collides with the user's own browser sessions.

• Package: @playwright/mcp@latest (npx, with --user-data-dir)
• No credentials in config — logins are per-site, stored in the profile cookies.

To log into a site for the first time:

~/claude_subagents/scripts/browser-profile-setup.sh https://dataegret.com/wp-admin/
# A Chromium window opens, you log in, you close it → cookies saved.
# Next agent invocation reuses the session.

Distinct from the official playwright plugin MCP (which uses an ephemeral profile with no saved logins). The browser-operator agent always uses playwright-persistent, never the ephemeral one.

---

The browser-operator sub-agent

A general-purpose browser automation agent, available from any project on any machine that has run install.sh. Drives the persistent Chromium profile above, so it never asks the user to grant Chrome extension permissions or log into sites twice.

Invoke it from any session for web-UI tasks that cannot be done via API:

Agent({
  subagent_type: "browser-operator",
  prompt: "Set the Yoast 'noindex' flag on https://dataegret.com/wp-admin/post.php?post=1196&action=edit
           (this is the 'Sorry to see you go!' page). Verify the front-end
           https://dataegret.com/news-blog/sorry-to-see-you-go/ shows
           'noindex, follow' in the robots meta tag after saving."
})

The agent uses the persistent Chromium profile, snapshots the page to orient, performs the action, and verifies the change via both the admin UI and a curl of the front-end. If a login is missing, it stops and tells you which URL to run through browser-profile-setup.sh.

---

Plugins

Install interactively after running install.sh:

claude plugin install php-lsp
claude plugin install frontend-design
claude plugin install skill-creator
claude plugin install playwright

---

SSH access (dataegret servers, requires YubiKey)

Add to ~/.ssh/config:

Host de-staging
  HostName svc.pgco.me
  Port 50222
  User ik

Host de-production
  HostName svc.pgco.me
  Port 50202
  User ik

Then test: ssh de-staging Touch YubiKey when prompted.

---

Project repo (dataegret.com website)

mkdir -p ~/work/de/website
cd ~/work/de/website
git clone git@github.com:hydrobiont/de-website-3.git de-website-3
cp de-website-3/.env.example de-website-3/.env
# Edit .env: add real Elementor Pro and ACF Pro license keys
docker compose -f de-website-3/docker-compose.yml up -d

Local dev runs at http://localhost:8080.

---

Keeping memory in sync

The memory/de-website/MEMORY.md file is the source of truth for project context. After significant work sessions, copy the updated memory back to the repo:

cp ~/.claude/projects/-Users-$(whoami)-work-de-website/memory/MEMORY.md \
   ~/claude_subagents/memory/de-website/MEMORY.md
git add memory/de-website/MEMORY.md
git commit -m "Update de-website project memory"
git push

On the other machine, pull and reinstall memory:

git pull
cp ~/claude_subagents/memory/de-website/MEMORY.md \
   ~/.claude/projects/-Users-$(whoami)-work-de-website/memory/MEMORY.md