Coding Agent — OpenClaw Skill | KiwiClaw Skills Hub

What This Skill Does

The Coding Agent skill turns your OpenClaw agent into an orchestrator of AI coding sub-agents. Instead of writing code line by line, your agent can spawn dedicated coding tools -- Codex (OpenAI), Claude Code (Anthropic), OpenCode, or Pi -- as background processes to handle complex programming tasks. These sub-agents can build new features, review pull requests, refactor large codebases, and perform iterative coding that requires file exploration and editing.

Each coding agent runs in a specified working directory with PTY (pseudo-terminal) support, which is required for these interactive terminal applications to function correctly. The skill supports both one-shot execution for quick tasks and background mode for longer operations, where you can monitor progress, send input, and receive completion notifications without blocking your main conversation.

The skill excels at parallelism. You can spawn multiple coding agents simultaneously -- for example, fixing several issues in parallel using git worktrees, or reviewing multiple PRs at once. Each agent works in its own isolated directory, and your OpenClaw agent coordinates the results. For PR reviews, the skill clones to a temp directory to keep your main repository untouched. The code execution use case page covers more workflows.

Example Prompts

Use Codex to build a REST API for a todo app with CRUD endpoints, SQLite database, and basic auth -- put it in ~/Projects/todo-api

Clone acme/webapp and review PR #130 -- tell me about code quality, potential bugs, and whether the tests are adequate

Refactor the authentication module in ~/Projects/myapp to use JWT tokens instead of session cookies

Fix issues #78 and #99 in parallel using git worktrees -- create a PR for each when done

Use Claude Code to add comprehensive error handling to all the API endpoints in ~/Projects/backend

Spawn a coding agent to write unit tests for the payment processing module and report back what coverage looks like

Build a dark mode toggle component for our React app in ~/Projects/frontend -- use Codex in full-auto mode

Requirements

Binary dependency: At least one of claude, codex, opencode, or pi

Codex: Install via npm or the OpenAI CLI. Default model is gpt-5.2-codex.
Claude Code: Install the Anthropic CLI. Requires Anthropic API key.
Pi: npm install -g @mariozechner/pi-coding-agent
Git: Required for Codex (it only runs inside git repositories)
PTY support: The bash tool must support pty:true for interactive terminal output

Setup on KiwiClaw

This skill is pre-installed and configured on all KiwiClaw plans. Codex and Claude Code are available on every tenant machine. On the Standard plan ($39/mo), LLM access for the coding sub-agents is included in your managed usage cap. On the BYOK plan ($15/mo), the agents use your own API keys. No setup needed.

Setup Self-Hosted

Install your preferred coding agent CLI (e.g., npm install -g @openai/codex)
Authenticate the CLI with your API keys
Ensure git is installed (Codex requires a git repository)
The skill will be automatically available once any supported coding agent binary is on the PATH

Related Skills

GitHub -- create PRs and manage issues for coding agent output
Slack -- get notified when coding tasks complete
Notion -- document architecture decisions and technical specs
Trello -- track coding tasks across your project board

FAQ

What coding agents does this skill support?

The Coding Agent skill supports Codex (OpenAI's CLI, using gpt-5.2-codex by default), Claude Code (Anthropic's coding agent), OpenCode, and Pi coding agent. Your OpenClaw agent can spawn any of these as background processes to handle complex coding tasks.

Can the coding agent build entire features from scratch?

Yes. You can ask it to build new features, create full applications, or scaffold entire projects. The agent spawns a coding sub-agent in the target directory with full-auto mode, which can create files, install dependencies, run tests, and commit changes.

Is it safe to let the coding agent modify my codebase?

The skill supports multiple safety levels. Default mode asks for approval before changes. Full-auto mode (--full-auto) auto-approves within a sandboxed workspace. The skill never modifies OpenClaw's own installation directory and always operates within the specified working directory. Read more about our skills vetting process.

Can the coding agent review pull requests?

Yes. The skill can clone a repo to a temp directory, check out a PR branch, and spawn a coding agent to review the diff against the base branch. This keeps your main repository untouched while providing thorough code review.