Two years of daily work with AI coding agents, distilled into a reproducible system. This repo is the full inventory: what's installed, how the pieces connect, and the patterns that work. It also serves as the portable source of truth for reproducing the setup on a new machine.

The Setup

Harness	Role	MCP Servers
Claude Code	Primary. Plugins, skills, hooks, commands all live here.	10 servers: Exa, Google-PSE, Granola, HelpScout, Slack, nanobanana, Docker MCP, ChatPRD, 2 plugin MCPs
Codex CLI	Secondary. Gets skills via symlink.	Docker MCP, Playwright, Pencil, Linear
Gemini CLI	Tertiary. Gets skills via symlink.	Docker MCP, Playwright
OpenCode	Fourth. Gets skills via symlink.	Docker MCP, 4 Z.AI servers

By the Numbers

	Count	What
🏪	10	Plugin marketplaces (Anthropic official, community, and personal)
🔌	65	Plugins providing agents, hooks, LSP integrations, and document tools
⚡	77	Unique skills (10 personal + 67 from plugins), synced across all harnesses
🌐	10+	MCP servers configured per-harness, Docker MCP Toolkit shared across all four
🔧	30+	CLI tools for service integrations, browser automation, and development runtimes
🖥️	20+	Apps and infrastructure including Docker Desktop, Obsidian, Superwhisper, and supporting macOS apps

Workflows

Patterns for how the tools compose together. These are atomic building blocks, not rigid pipelines. See workflows/README.md for details.

Pattern	What It Does
Multi-Model Consultation	Three models (Gemini, Codex, Claude) review every hard problem in parallel via the braintrust plugin
Local Quality Gate	Run the quality-gate skill with --local during development for early issue detection
PR Review Pipeline	Full review combining agent reviews, Devin Review, and CodeRabbit in parallel
Design & Prototyping	Agentic CLI + frontend-design skill + nanobanana, or external tools (Stitch, v0, Lovable, Paper, Pencil)
Security	Layered: pre-commit credential scanning, codeguard-security plugin, full PR review pipeline
Knowledge Capture	Obsidian Knowledge Vault with Zettelkasten + CORE framework, semantic search via obsidian-graph-mcp (Voyage Context-3 + pgvector)
Skill Distribution	Automatic. Plugins install skills into Claude Code, sync script symlinks to all other harnesses on every session start

Configuration Reference

Per-harness configuration, global instructions, hooks, and security settings.

Config	What
Claude Code	CLAUDE.md global instructions, settings, hooks
Codex CLI	Settings, MCP servers, trusted projects
Gemini CLI	Settings, MCP servers, global instructions
OpenCode	Settings, MCP servers
Git	Global pre-commit hook (credential blocking, .env warning, npm audit)

References

Reference	What
Browser tools comparison	Command-by-command matrix: agent-browser (~130) vs dev-browser (~79) vs Claude-in-Chrome (~39) vs Playwright MCP (~42)

---

How Skill Sync Works

The only executable code in this repo. Claude Code is the single source of truth for plugins and skills. The other three harnesses get read access through symlinks.

Claude Code plugins install skills
        |
        v
  sync-skills.sh scans ~/.claude/skills/ and ~/.claude/plugins/cache/
        |
        v
  Deduplicates by name (personal skills win over plugin skills)
        |
        v
  Creates symlinks in ~/.gemini/skills/
                       ~/.codex/skills/
                       ~/.config/opencode/skills/

The script runs automatically via a SessionStart hook every time Claude Code starts. Silent, failure-safe.

zsh scripts/sync-skills.sh           # sync all skills
zsh scripts/sync-skills.sh --dry-run # preview without changes

Reproducing on a new machine

1. Install the harnesses

curl -fsSL https://claude.ai/install.sh | bash          # Claude Code
npm install -g @google/gemini-cli                        # Gemini CLI
brew install --cask codex                                # Codex CLI
curl -fsSL https://opencode.ai/install | bash            # OpenCode

2. Clone this repo

git clone https://github.com/drewburchfield/agentic-setup.git ~/dev/agentic-setup

3. Wire the sync script

mkdir -p ~/.claude/scripts
ln -s ~/dev/agentic-setup/scripts/sync-skills.sh ~/.claude/scripts/sync-skills.sh

Then add the SessionStart hook to ~/.claude/settings.json (see hooks.md for the exact JSON).

4. Install plugins

See plugins.md for the full list. Marketplaces are added with claude plugin marketplace add <url>, then individual plugins are enabled in settings.

5. Configure MCP servers

See mcp-servers.md for per-harness configuration. Docker MCP Toolkit is shared across all four harnesses.

6. Copy global instructions and references

cp ~/dev/agentic-setup/config/claude-code/CLAUDE.md ~/.claude/CLAUDE.md
mkdir -p ~/.claude/references
cp ~/dev/agentic-setup/references/* ~/.claude/references/

See the settings reference for permissions, environment variables, and other settings to configure.

Day-to-day operations

Adding a personal skill

All skills go through Claude Code. Never use npx skills add in project directories; it creates per-agent scaffolding that conflicts with the sync script.

cp -r /path/to/skill-name ~/.claude/skills/skill-name    # must contain SKILL.md
zsh ~/dev/agentic-setup/scripts/sync-skills.sh            # or wait for next session start

For npm packages, symlink instead of copying:

ln -s /opt/homebrew/lib/node_modules/agent-browser/skills/agent-browser ~/.claude/skills/agent-browser

Adding a plugin

claude plugin marketplace add <github-url>    # add a new marketplace
# Then enable individual plugins from the marketplace

Adding an MCP server

MCP servers are configured per-harness. See mcp-servers.md for config file locations and format.

What NOT to do

• Don't use npx skills add in project directories. It scatters .agent/, .claude/, .goose/, .kiro/, .windsurf/, and skills/ directories that duplicate what the sync script handles.
• Don't install skills directly into secondary harnesses. The sync script owns those directories.
• Don't use Gemini's Antigravity skill installer. It creates a separate index that duplicates what the sync script provides.