Terminal-first coding agent tuned for Claude workflows.
Demo highlights:
- Slash command discovery: Type
/to browse locally-scoped prompts with Tab completion - Natural argument passing: Use commands like
/tell_me_a_joke_about gpt-5-codexwith Claude Code-style argument parsing
Note
Claude Codex extends the upstream openai/codex project. Follow the upstream repository for the canonical feature list; this README focuses on the extras maintained in this fork.
- Rust toolchain with
cargo(install via https://rustup.rs/). - Optional:
mpg123to play success sounds after local installs.
git clone https://github.com/anthropics/claude-codex.git
cd claude-codex
./install-local.shThe script builds the Rust CLI in release mode and installs the resulting codex binary into ~/bin. Re-run ./install-local.sh anytime you pull new changes.
Launch the agent with:
codex- Project-scoped slash prompts. Codex auto-discovers
.codex/prompts/inside your Git repo, hoists nested folders as/folder:promptcommands, and merges them with your global prompt library. Optional frontmatter fields (description,argument-hint) power richer TUI tooltips and argument prompts. Seedocs/prompts.md. - Prompt arguments with structured metadata. Provide arguments right after a slash command (for example
/review api.rs bugs). The composer injectsargument_n:headers plus the prompt's metadata before sending it to the agent. - Native stop hooks for automation. Configure
[hooks.stop.*]tables in.codex/config.tomlto launch scripts when a turn finishes. Hooks receive JSON on stdin, expand$CODEX_PROJECT_DIRinside commands/args/env, and respect per-hook timeouts. Checkcodex-rs/example-configs/hooks-config.tomland.strategic-claude-basic/core/hooks/stop-session-notify.pyfor templates. - Strategic Claude workflows. The
.strategic-claude-basic/directory documents research notes, plans, validation scripts, and reusable hooks tailored for Claude-oriented development loops. - Local install script & audio cues.
./install-local.shcompiles the Rust CLI, installs it into~/bin, and optionally plays a celebratory clip fromassets/sounds/whenmpg123is available.
- Authentication: The agent still supports ChatGPT sign-in and API-key flows; see
docs/authentication.mdfor details. - Config files: Preferences live in
~/.codex/config.toml. Claude-specific additions (prompts, hooks) require no extra flags—drop files under.codex/and the CLI discovers them on the next session. - Model Context Protocol: Enable MCP integrations by defining
mcp_serversentries as described indocs/advanced.md#model-context-protocol-mcp.
For baseline CLI behavior, sandboxing guarantees, and release binaries, consult the upstream README: https://github.com/openai/codex. This fork intentionally omits npm/Homebrew distribution instructions in favor of the local install flow above.
docs/getting-started.md– General CLI usage walkthroughs.docs/prompts.md– Custom prompt layout, namespaces, and argument hints.docs/config.md– Full configuration reference, including hook tables.docs/sandbox.md– Approval modes, sandbox behavior, and environment variables.docs/advanced.md– Advanced workflows (CI mode, verbose logging, MCP).codex-rs/example-configs/hooks-config.toml– Hook configuration examples..strategic-claude-basic/– Plans, research, and scripts supporting Claude automation.
This repository is licensed under the Apache-2.0 License.
