Mastering Codex: A Complete Beginner‑to‑Advanced Configuration Guide

After testing Codex, the author finds its backend capabilities more stable and less error‑prone than Claude Code, though its frontend performance is slightly weaker. If you can tolerate a slower response, Codex is recommended because its subscription costs only $20 per month, and even lower prices are possible via proxy services.

Codex is OpenAI’s AI programming assistant built on the GPT‑5 model. By 2025 it has evolved from a simple code‑completion tool into a full‑stack assistant capable of writing code, fixing bugs, manipulating the file system, running commands, and executing tests. It runs in terminals, IDEs, cloud environments, GitHub, and even on mobile devices, with further enhancements after integrating the MCP protocol.

Installation begins by ensuring Node.js 22+ is installed on macOS, Linux, or Windows. The simplest method is a single npm command after Node.js is set up: npm install -g @openai/codex@latest On Windows, the author advises running the command prompt as an administrator to avoid permission errors.

After installation, verify the version with codex --version. The process typically takes about five minutes, considerably faster than configuring other AI coding tools.

If network access is limited, switch npm to a domestic mirror:

npm config set registry https://registry.npmmirror.com

Codex supports two authentication methods: an OpenAI API key (set via environment variables or system config files) and direct login with a ChatGPT Plus account using codex login. The author prefers the API‑Key approach for stability.

Key configuration parameters reside in ~/.codex/config.toml. Important settings include: model_reasoning_effort = "high" – improves code quality. disable_response_storage = true – prevents storing conversation history. approval_policy = "never" – lets Codex execute actions automatically (new users may start with ask to avoid accidental operations). sandbox_mode = "danger-full-access" – grants full file‑system access (used after gaining confidence).

A sample configuration file is provided, showing model providers (e.g., duckcoding and packycode), MCP router command settings, and tool toggles.

# Codex system prompt
model_provider = "duckcoding"
model = "gpt-5-codex"
model_reasoning_effort = "high"
disable_response_storage = true
approval_policy = "never"
sandbox_mode = "danger-full-access"
# ... (additional sections omitted for brevity)

Common Codex commands include: codex – starts an interactive session. /mode auto-edit – enables automatic file edits with confirmation. /mode full-auto – lets the AI take full control (recommended for experienced users). /mode suggest – default mode requiring user confirmation for each step. /model gpt-5-codex – switches to the full‑power model. /approvals auto – allows free operation within the current directory; full-access extends this to external resources. codex --resume – restores a previous session. codex exec "..." – runs non‑interactive tasks such as generating unit tests. codex -i screenshot.png "fix this error" – processes an image directly.

The Model Context Protocol (MCP) introduced at the end of 2024 standardizes communication between Codex and external tools. The author uses MCP Router to manage multiple servers, each offering specialized capabilities: context7 – fetches up‑to‑date technical documentation for new frameworks. zhipu-search – a Chinese‑optimized search engine for better AI comprehension. playwright – browser automation for web testing and crawling. duckduckgo – privacy‑preserving web search without an API key. sequential-thinking – assists with complex logical reasoning. serena – code‑semantic retrieval and symbol analysis. shrimp-task-manager – task‑flow management.

By configuring these MCP servers, Codex transforms from a simple code generator into a comprehensive AI development partner that can retrieve documentation, search the web, control browsers, and manage tasks.

The author also emphasizes the importance of the AGENTS.md file (located in ~/.codex/ or project‑specific directories) which defines the AI’s role, principles (KISS, YAGNI, SOLID, DRY), and workflow stages (understanding, planning, execution, reporting). A sample AGENTS.md is included, outlining roles such as "Technical Architect" and "Full‑Stack Expert" and prescribing engineering best practices.

Overall, the configuration turns Codex into an AI assistant that knows when to use which tool, how to handle exceptions, and how to maintain code quality throughout the software development lifecycle.