My Codex CLI default config ( ~/.codex )

AGENTS.md

Default Instructions for Codex CLI

• You have useful tools available as MCP or command-line. Use them.
• Unless you are absolutely sure that you have correct and, crucially, up-to-date information in your knowledge, always get information from the web. You can use Tavily for searching and getting information, and you can always use curl to fetch web pages.
• If you are working on something that presents web pages, you should use Playwright to open these pages, take snapshots, and inspect them.
• When working in a git repository, always switch to a new branch, unless explicitly insutrcted not to.
• If the git repo has a corresponding repo on GitHub, use the gh tool for things like looking at issues, opening pull requests, reading review comments, and looking at CI results.
• If the project has tests you can run locally, always run them and make sure everything works correctly.
• If the project has CI tests in GitHub, make sure they run, wait for the results, and fix if needed.

Source Control (Git)

• when invoking git commit, always use --author="Codex CLI"

Useful Command-Line Tools

GitHub

• Use the gh command-line to interact with GitHub.

JSON

• Use the jq command to read and extract information from JSON files.

RipGrep

• The rg (ripgrep) command is available for fast searches in text files.

Clipboard

• Pipe content into pbcopy to copy it into the clipboard. Example: echo "hello" | pbcopy.
• Pipe from pbpaste to get the contents of the clipboard. Example: pbpaste > fromclipboard.txt.

Web (HTTP/S)

• Use curl to fetch web pages.

MCP Tools

playwright

• Operate a web browser: navigate, click around, take screenshots
• Use this whenever you want to look at a web page. For example, when working on a web app, you can run the local web server and interact with the web app.
• This is especially important if you're working on anything visual - you should always take a look using playwright so that you know how the results of your work look.

markitdown

• Use this to convert various file formats to markdown.
• Very useful if you need to read files that are not supported natively by the model.

tavily

• Use this to run web searches. It is always better to search the web than to rely on your own knowledge, which may be outdated.
• You can also retrieve content in a format easy for ingestion. Use that if needed, but you can also just use curl if you have a URL.

thinking

• If you are not a "reasoning" model, trained to produce chanis-of-thought, you MUST use this tool to think through problems.
• If you are GPT-4.1, ALWAYS use the thinking tool to think through a problem.
• Always report back what you thought using the tool.
• If you are a reasoning model like o3, o4-mini, or codex-mini, you don't need to use this tool.

Python

• Unless instructed otherwise, always use the uv Python environment and package manager for Python.
  • uv run ... for running a python script.
  • uvx ... for running program directly from a PyPI package.
  • uv ... ... for managing environments, installing packages, etc...

JavaScript / TypeScript

• Unless instructed otherwise, always use deno to run .js or .ts scripts.
• Use npx for running commands directly from npm packages.

Documentation Sources

• If working with a new library or tool, consider looking for its documentation from its website, GitHub project, or the relevant llms.txt.
  • It is always better to have accurate, up-to-date documentation at your disposal, rather than relying on your pre-trained knowledge.
• You can search the following directories for llms.txt collections for many projects:
  • https://llmstxt.site/
  • https://directory.llmstxt.cloud/
• If you find a relevant llms.txt file, follow the links until you have access to the complete documentation.

config.toml

approval_policy = "never"
sandbox_mode = "danger-full-access"

file_opener = "vscode-insiders"

profile = "o3"

[profiles.o3]
model = "o3"
model_reasoning_effort = "high"
model_reasoning_summary = "detailed"

[profiles.gpt]
model = "gpt-4.1"

[mcp_servers.playwright]
command = "npx"
args = ["-y", "@playwright/mcp@latest"]

[mcp_servers.markitdown]
command = "uvx"
args = ["markitdown-mcp"]

[mcp_servers.tavily]
command = "npx"
args = ["-y", "tavily-mcp"]
env = { "TAVILY_API_KEY" = "tvly-XxXxXxXxXxXxXXXxXxXx" }

[mcp_servers.thinking]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-sequential-thinking"]