Reusable tools, prompts, instructions, skills, and hooks for coding agents. Shared components live alongside explicit integrations for individual harnesses; each component documents where it works.
| Component | Harness | Documentation |
|---|---|---|
| Session history and resume | Claude Code | Utilities |
| Session history and resume | Codex | Codex integration |
| Subscription and API authentication profiles | Codex | Authentication profiles |
| Personal context, compaction, and Vim preferences | Codex; bootstrap reference | Configuration template |
| Keyring-backed API key wrapper | Codex | Store the key in the keyring |
| Tmux launcher with backend and status display | Codex | Tmux status display |
| Memory export, import, and status | Claude Code | Utilities |
| Release digest against local configuration | Claude Code | Utilities |
| Release digest against local configuration | Codex | Release digest |
| Statusline and quota display | Claude Code | Settings |
| Hooks | Claude Code | Hooks |
| Teaching skill | Claude Code and Codex | Shared workflow |
| Baseline adoption skill | Claude Code and Codex | Shared workflow |
| Release digest skill | Claude Code and Codex | Shared workflow |
| Review and browser verification prompts | Any harness accepting text prompts | Prompts |
| Optional instruction modules | Any harness accepting instruction files | Instructions |
| Hook reference examples | Claude Code payloads; not deployed | Examples |
| Agent configuration examples | Claude Code | Agents |
Choose a harness and scope, then preview the selected components:
git clone https://github.com/adereis/agent-toolbox.git
cd agent-toolbox
python3 tools/install.py --harness codex --scope user \
--component skills --component scriptsAdd --apply to create the links. Existing conflicting files are preserved.
See installation for scope, components, configuration,
and removal. /sync in Claude Code wraps this same installer. Session
utilities run from a terminal; --list and --json only inspect history.
Most people install and operate these components through a coding agent rather than by hand. Point the agent at this repository and describe the outcome you want; it reads the component table above, runs the installer, and follows the README for the harness it selected.
Components are written for that workflow. When a component cannot run on your platform it exits with the command that works there instead of failing quietly, so the agent can finish the task without a round trip through you. Supported platforms are stated with each component; Linux is tested, and macOS support varies by component.
agent-toolbox/
├── skills/ # Shared workflows
├── prompts/ # Optional task prompts
├── instructions/ # Opt-in policy modules
├── tools/ # Installer and common utility code
├── harnesses/
│ ├── claude-code/ # Claude wrappers, hooks, settings, utilities
│ └── codex/ # Codex profiles, wrappers, and session utility
├── examples/ # Reference implementations, excluded from install
├── docs/ # Installation and migration
├── tests/ # Automated and live integration checks
├── AGENTS.md # Repository development conventions
├── CLAUDE.md # Claude entry point to those conventions
└── README.md # Catalog and installation
The GitHub repository was renamed from claude-code-extensions to
agent-toolbox, retaining its
history. The claude-code-baseline tag marks
the collection before this migration. Source directories formerly at the
repository root now live under harnesses/claude-code/; update any links or
scripts that refer to checkout paths. Installed ~/.claude/ paths are
unchanged by this source reorganization. Follow the
migration notes to remove retired installations and
update session utility links.
- mcp-servers: reusable MCP servers.
- claude-sandbox: a containerized Claude Code environment.
MIT