claude-in-a-box

claude-in-a-box

Dotfiles-style, reproducible Claude Code setup — my whole stack, packaged. Agentic-install only: there is no monolithic install script. Point an agent at INSTALL.md and it installs each tool the way its own maintainers document (brew/curl/cargo, claude plugin, rtk init, ...), adapting to whatever OS/package manager it finds, and calls out to a few small helper scripts only for the fiddly file operations (symlink+backup, JSON settings merge, docker compose wiring).

Bar is set to oh-my-claudecode. CLI + skills preferred over MCP; MCP is used only where there's no good CLI/skill path.

Installation

Claude stack

For humans

Paste this prompt to your LLM agent (Claude Code, etc.):

Set up this machine with my Claude Code stack by following the guide here:
https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL.md

You can also follow INSTALL.md yourself — it's plain shell commands with reasoning attached, not a script you must run as a black box — but letting an agent drive means the interactive parts (API keys, verification, adapting commands to your actual OS) get handled for you.

For LLM agents

Fetch the guide and follow it:

curl -s https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL.md

Codex stack

For humans

Install the Claude stack first (shared services — jina, OTel, claude-mem — run once for both), then:

Set up this machine with my Codex stack by following the guide here:
https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL_CODEX.md

For LLM agents

curl -s https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL_CODEX.md

Antigravity stack

For humans

Install the Claude stack first (shared services - jina, OTel, claude-mem), then:

Set up this machine with my Antigravity stack by following the guide here:
https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL_ANTIGRAVITY.md

For LLM agents

curl -s https://raw.githubusercontent.com/ZebaLive/claude-in-a-box/main/INSTALL_ANTIGRAVITY.md

What it installs

Claude stack

GitHub is handled by the gh CLI (no MCP server).

Codex stack

Shares jina, OTel, and the claude-mem database with the Claude stack. All four Claude plugins now have Codex equivalents.

Why these transports

• Skills/CLI first (ponytail principle): fewer moving parts, less token overhead than MCP.
• jina = skill, not MCP. Web fetch/screenshots stay on-machine via a self-hosted Reader. The skill is plain curl to localhost:3333. See jina-ai/README.md.
• exa = MCP. Web search needs a cloud index; no local equivalent.
• context7 = CLI. Live library docs via npx ctx7 library / ctx7 docs, driven by rules/context7.md. No MCP server, no token overhead when unused. CONTEXT7_API_KEY optional (rate limits only).
• github = gh CLI. No MCP server — gh covers PRs, issues, releases, and gh api.
• rtk = CLI + hook, not MCP. A PreToolUse hook transparently rewrites Bash commands (git status, cargo test, ...) to compact rtk equivalents before execution — 60-90% less token usage per command, zero prompting overhead.

Secrets

Keys live in .env (gitignored) and are read by the agent at install time (set -a && . ./.env && set +a). Nothing secret is committed. This repo is safe to make public. Rotate the keys that were ever pasted into a local ~/.claude.json before publishing.

Layout

INSTALL.md              # Claude stack entrypoint — step-by-step agent instructions
INSTALL_CODEX.md        # Codex stack entrypoint — mirrors INSTALL.md for Codex CLI
INSTALL_ANTIGRAVITY.md  # Antigravity stack entrypoint - source-verified plugin/CLI wiring
.env.example            # required API keys
claude/                 # shared Claude config, symlinked into ~/.claude
  CLAUDE.md             # global instructions (OMC orchestration, routing, git)
  rules/context7.md     # Context7 docs-lookup rule
  link.sh               # helper: links CLAUDE-IN-A-BOX.md and imports it from ~/.claude/CLAUDE.md
  merge-settings.sh     # helper: merges shared-settings.json into ~/.claude/settings.json
codex/                  # shared Codex config, merged into ~/.codex
  AGENTS.md             # global instructions (OMX orchestration, shared memory)
  link.sh               # helper: upserts CODEX-IN-A-BOX block into ~/.codex/AGENTS.md
  merge-config.sh       # helper: merges OTel + sandbox defaults into ~/.codex/config.toml
  shared-config.toml    # base config (OTel grpc exporters, sandbox_mode)
antigravity/            # shared Antigravity prompt/config snippets
  GEMINI.md             # global instructions copied to ~/.gemini/GEMINI.md
skills/jina-reader/     # the jina-reader skill (SKILL.md)
jina-ai/                # local Reader stack (docker-compose; colima+LaunchAgent on
                         #   macOS, native Docker+systemd --user on Linux)
  setup-jina.sh         # helper: brings up the Reader, installs the login service
monitoring/             # OTel telemetry collector for Claude Code
  setup-otel.sh         # helper: brings up the collector, installs the login service

rtk itself lives outside this repo (rtk-ai/rtk). INSTALL.md has the agent install the binary (brew on macOS, vendor script on Linux) and run rtk init -g --auto-patch itself, rather than us vendoring a copy of the hook JSON it writes into ~/.claude/settings.json — that way it always matches whatever the installed rtk version actually expects.

This stack is opinionated: jina, OTel, and rtk always install — there are no skip flags. Docker (colima on macOS) is a hard requirement, same as the Claude Code CLI itself.

Shared config vs. machine-local

claude/CLAUDE.md and claude/rules/ are symlinked into ~/.claude, so edits in this repo take effect live and are versioned. link.sh backs up any existing real file to .bak before linking.

Machine-specific instructions (telemetry endpoints, SSH hosts, cloud profiles) belong in ~/.claude/CLAUDE.local.md — not managed here. CLAUDE.md has a commented @CLAUDE.local.md import; uncomment it to load them.

Ordering gotcha, handled in INSTALL.md: both omc setup and rtk init want to write ~/.claude/CLAUDE.md themselves (OMC refuses outright if it's already our symlink; rtk isn't as careful and will write straight through the symlink into the repo's tracked file). INSTALL.md runs omc setup and rtk init before link.sh, so they land on a real file, then link.sh backs that generated file up and puts our symlink back on top. Our CLAUDE.md carries a permanent @RTK.md import so the rtk-generated meta-command reference (rtk gain, rtk discover, ...) still loads even though the line rtk itself appends gets overwritten.