scaffold-project

AI Tools

Claude Code plugins and sandboxed Docker environments for AI coding assistants.

Claude Code Plugins

Install via the Claude Code marketplace:

/plugin marketplace add ShravanSunder/ai-tools

Then install individual plugins with /plugin install <name>@ai-tools.

Plugin	Description
ai-scaffold	Project scaffolding with standard dev configs (biome, ruff, vitest, pytest, cursor rules, claude hooks)
skill-peekaboo	Visual UI testing for macOS apps using Peekaboo CLI
quorum-counsel	Multi-model review orchestration -- counsel-reviewer and codex-solver background agents

See plugins/ for full details.

---

Agent Sidecar

Run AI coding agents (Claude Code, Codex, Gemini CLI) inside Docker containers with network-level isolation. The agent gets full workspace access but can only reach domains you explicitly allow.

# From any git repository
run-agent-sidecar.sh --run-claude

Why a sidecar container?

AI agents with tool use can execute arbitrary shell commands, install packages, and make network requests. Running them in a container with an egress firewall provides a practical security boundary without giving up functionality:

flowchart LR
    subgraph host ["Host Machine"]
        ctl["sidecar-ctl.sh\n(firewall control)"]
        run["run-agent-sidecar.sh"]
    end

    subgraph container ["Docker Container"]
        fw["iptables + dnsmasq\n(egress firewall)"]
        agent["AI Agent\n(claude / codex / gemini)"]
        ws["/workspace\n(bind mount)"]
    end

    run -->|"build + start"| container
    ctl -->|"allow / block domains"| fw
    agent -->|"read/write files"| ws

    agent -->|"network requests"| fw
    fw -->|"allowed domains only"| internet["npm, pypi, AI APIs\n(allowlisted)"]
    fw -.->|"blocked"| blocked["everything else"]

Concern	How it's handled
Arbitrary network access	Egress firewall blocks all traffic except allowlisted domains (npm, pypi, AI APIs). Toggle presets for GitHub push, Notion, Linear.
Persistent state leaking	Named volumes isolate shell history, venvs, node_modules per workspace. Container recreation is cheap (--reload ~5s).
Package supply chain	APT repos blocked at firewall level after build. Runtime installs impossible.
Git corruption	.git/ mounted read-only. Agent can read history but cannot rewrite refs or force-push.
Config tampering	.agent_sidecar/ shadowed with empty tmpfs. Agent cannot read or modify sidecar configuration.
Scope creep across repos	Each repo gets its own container, volumes, and firewall rules. No cross-repo contamination.

Configuration at a glance

Three-tier config hierarchy lets you customize per-team and per-developer without forking:

Base (agent_sidecar/setup/)     -- defaults shipped with this repo
  + Repo (.agent_sidecar/*.repo.*)  -- team overrides, committed
    + Local (.agent_sidecar/*.local.*) -- personal overrides, gitignored

Additive files (firewall allowlists, zshrc, init scripts) merge all tiers. Override files (config, dockerfile) pick the highest-priority tier.

Quick reference

run-agent-sidecar.sh --run-claude    # Start Claude Code in sidecar
run-agent-sidecar.sh --run-codex     # Start Codex
run-agent-sidecar.sh --run-gemini    # Start Gemini CLI
run-agent-sidecar.sh --reload        # Recreate container (~5s)
run-agent-sidecar.sh --full-reset    # Rebuild image + recreate (~2-5min)

sidecar-ctl firewall allow notion    # Allow Notion API
sidecar-ctl firewall toggle 15m      # Enable all presets for 15 minutes
sidecar-ctl firewall clear           # Revoke all toggle access
sidecar-ctl status                   # Show container + firewall state

See agent_sidecar/ for full setup, architecture, and configuration docs.

---

Repository Structure

ai-tools/
├── plugins/                     # Claude Code plugins
│   ├── ai-scaffold/             # Project scaffolding
│   ├── skill-peekaboo/          # macOS visual UI testing
│   └── quorum-counsel/          # Multi-model review orchestration
├── skills/                      # Pure skills (future)
├── agent_sidecar/               # Docker sidecar system
└── CLAUDE.md                    # Agent instructions