Back to All Cheatsheet Libraries
cheatsheets
Model and API parameter reference plus setup guidance for building with OpenAI's Codex/code-focused models.
# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"
# Package managers
npm install -g @openai/codex
brew install --cask codex
codex # start it
codex app # launch the desktop app
Sign in with your ChatGPT plan (Plus, Pro, Business, Edu, Enterprise) — that is the recommended path and needs no API key. An API key works too but requires extra setup and bills separately.
| Subcommand | Purpose |
|---|---|
codex | Interactive TUI session. The default. |
codex exec "task" | Non-interactive. The scriptable surface — run a task, get output, exit. This is what you use in CI. |
codex review | Run a code review non-interactively. |
codex apply | Apply the agent's latest diff as a git apply. |
codex resume / fork | Resume or branch a previous session — picker by default. |
codex queue | Queue a message for an existing session. |
codex archive / unarchive / delete | Session lifecycle management. |
codex agents | Browse all agent sessions on the shared local daemon. |
codex sandbox | Run arbitrary commands inside Codex's sandbox — useful on its own. |
codex mcp | Manage external MCP servers. |
codex mcp-server | Runs Codex itself as an MCP server over stdio, so another agent can drive it. |
codex plugin | Manage Codex plugins. |
codex login / logout | Authentication. |
codex doctor | Diagnose the install and update to the latest version. |
codex completion | Generate shell completion scripts. |
codex cloud | Experimental — browse Codex Cloud tasks and apply changes locally. |
codex execpolicy | Execution-policy tooling. |
codex features | Inspect feature flags. |
AGENTS.md and configuration
AGENTS.mdis Codex's project-instruction file — the equivalent ofCLAUDE.md. Generate one with/init.- It is an open, cross-tool convention, not an OpenAI-only format — several agents read it.
- Config lives in
~/.codex/config.toml. Admins can enforce policy throughrequirements.toml, includingallow_managed_hooks_onlyto ignore user and project hook configs. /importpulls your setup, project config and recent chats across from Claude Code — a genuinely useful migration path in either direction.
| Group | Command | What it does |
|---|---|---|
| Session | /model | choose what model and reasoning effort to use |
| Setup | /ide | — |
| Security | /permissions | choose what Codex is allowed to do |
| Setup | /keymap | remap TUI shortcuts |
| Setup | /vim | toggle Vim mode for the composer |
| Security | /setup-default-sandbox | set up elevated agent sandbox |
| Security | /sandbox-add-read-dir | — |
| Setup | /experimental | toggle experimental features |
| Security | /approve | approve one retry of a recent auto-review denial |
| Context | /memories | configure memory use and generation |
| Setup | /skills | use skills to improve how Codex performs specific tasks |
| Setup | /import | import setup, this project, and recent chats from Claude Code |
| Setup | /hooks | view and manage lifecycle hooks |
| Code | /review | review my current changes and find issues |
| Session | /rename | rename the current thread |
| Session | /new | start a new chat during a conversation |
| Session | /archive | archive this session and exit |
| Session | /delete | permanently delete this session and exit |
| Session | /resume | resume a saved chat |
| Session | /fork | fork the current chat |
| Surfaces | /app | continue this session in the Desktop app |
| Context | /init | create an AGENTS.md file with instructions for Codex |
| Context | /compact | summarize conversation to prevent hitting the context limit |
| Session | /plan | switch to Plan mode |
| Session | /goal | set or view the goal for a long-running task |
| Agents | /agents | view and switch between all active agent sessions |
| Agents | /side | — |
| Session | /btw | — |
| Output | /copy | copy the last response, code block, or quote |
| Output | /export | export the conversation as markdown |
| Output | /raw | toggle raw scrollback mode for copy-friendly terminal selection |
| Code | /diff | show git diff (including untracked files) |
| Context | /mention | mention a file |
| Info | /status | show current session configuration and token usage |
| Files | /cd | change the current working directory |
| Files | /pwd | show the current working directory |
| Info | /usage | view account usage or use a usage limit reset |
| Info | /debug-config | show config layers and requirement sources for debugging |
| Setup | /title | configure which items appear in the terminal title |
| Setup | /statusline | configure which items appear in the status line |
| Setup | /theme | choose a syntax highlighting theme |
| Extras | /pets | choose or hide the terminal pet |
| Integrations | /mcp | list configured MCP tools; use /mcp verbose for details |
| Integrations | /apps | manage apps |
| Integrations | /plugins | browse plugins |
| Account | /logout | log out of Codex |
| Session | /quit | — |
| Session | /exit | exit Codex |
| Info | /feedback | send logs to maintainers |
| Info | /rollout | print the rollout file path |
| Processes | /ps | list background terminals |
| Processes | /stop | stop all background terminals |
| Session | /clear | clear the terminal and start a new chat |
| Setup | /personality | choose a communication style for Codex |
| Agents | /subagents | switch between this session's subagents |
# The Responses API is the current primary interface
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY
resp = client.responses.create(
model="gpt-5.6",
input="Summarise this changelog in three bullets.",
)
print(resp.output_text)
# Chat Completions still works and is widely supported
resp = client.chat.completions.create(
model="gpt-5.6",
messages=[{"role": "user", "content": "Hello"}],
)
Generating images
# Standalone — Image API
result = client.images.generate(
model="gpt-image-2",
prompt="A cutaway diagram of a server rack, technical illustration",
size="1536x1024",
quality="high",
output_format="png", # png (default) | jpeg | webp
background="transparent", # transparent | opaque | auto
)
# Returns base64-encoded image data
# In-conversation — Responses API tool
resp = client.responses.create(
model="gpt-5.6",
input="Draw the architecture we just discussed, then label the cache layer.",
tools=[{"type": "image_generation"}],
)
| Model | Notes |
|---|---|
| gpt-image-2 | Current flagship. Accepts custom resolutions, not just fixed presets. |
| gpt-image-1.5 | Previous generation. |
| gpt-image-1 | Older. |
| gpt-image-1-mini | Smaller and cheaper. |
Size constraints that will reject your request
- Common sizes:
1024x1024,1536x1024,1024x1536,2048x2048,3840x2160. Default isauto. - Max edge 3840px. Dimensions must be multiples of 16.
- Aspect ratio capped at 3:1 — no ultra-wide banners in one shot.
- Total pixels must fall between 655,360 and 8,294,400.
quality:low/medium/high/auto.output_compression(0–100) applies to JPEG and WebP only.- Output is base64-encoded in the response — you decode and write the file yourself.
PDFs, ZIPs and everything else
# Code Interpreter — works with Responses, Chat Completions and Assistants
container = client.containers.create(name="report-box", memory_limit="4g")
resp = client.responses.create(
model="gpt-5.6",
tools=[{"type": "code_interpreter", "container": container.id}],
input="Read the attached CSV, chart revenue by month, "
"and save a PDF report plus a ZIP of the source charts.",
)
# Generated files come back as annotations on the response:
# container_file_citation -> { container_id, file_id, filename }
# Retrieve each with the "Retrieve container file content" endpoint.
Container limits worth knowing before you build on it
- A container expires after 20 minutes idle, and expired containers are inaccessible. Download generated files before then — this is the single most common way people lose output.
- Memory tiers: 1 GB (default), 4 GB, 16 GB, 64 GB. Set
memory_limitat creation for anything data-heavy. - Rate limit: 100 requests per minute per organisation.
- Files are retrieved by
container_file_citationannotation, not returned inline — you must make a second call per file. - Inputs it accepts: CSV, JSON, PDF, XLSX, DOCX, PPTX, images, source code, ZIP and TAR archives.
| How | Formats | Needs |
|---|---|---|
| Written directly | .md .txt .csv .tsv .json .yaml .toml .xml .html .svg .css .js .ts .py .rs .go .sh .sql .ipynb .env and any other plain text | Nothing. No code executed, no library required. |
| Script | .pdf | reportlab, weasyprint, or render HTML → PDF. Merging and form-filling via pypdf. |
| Script | .xlsx .xlsm | openpyxl for formulas and formatting, pandas for bulk data. |
| Script | .docx .pptx | python-docx, python-pptx, or the JS equivalents. |
| Script | .zip .tar .tar.gz .gz | Shell zip/tar, or Python zipfile/tarfile. Nothing to install. |
| Script | .png .jpg .webp .gif | matplotlib / plotly for charts, Pillow for manipulation. Rendered from code, not drawn. |
| Script | .parquet .sqlite .db .feather | pandas, pyarrow, sqlite3. |
| API call | .png .jpeg .webp from a prompt | An OpenAI API key and gpt-image-2. The only format generated by a model rather than by code. |
The limits that actually bite
- Context, not file size, is the constraint. Reading a large CSV into the conversation wastes the window and degrades the answer. Tell Codex to process it with pandas instead — the file then never enters context and size stops mattering.
- The sandbox governs where files can be written. Codex confines writes by default;
/permissionsand/setup-default-sandboxcontrol it. A script that writes outside the workspace will be blocked, not silently fail. - XLSX caps at 1,048,576 rows × 16,384 columns per sheet — an Excel format limit, not a Codex one. Beyond that, split sheets or use Parquet.
- Codex cannot see the rendered output. It writes plotting code without looking at the chart, so clipped labels and unreadable colour choices go unnoticed unless you check.
- Missing libraries are a permission prompt, not magic — Codex will offer to install them.
Gotchas
Worth knowing
- The old Codex model is dead. Any tutorial referencing
code-davinci-002is describing a product retired in 2023 and unrelated to this one. - Container files vanish after 20 minutes idle. If you build a pipeline on Code Interpreter, download before the container expires or the output is gone.
- Images return as base64, not a URL. You decode and save them yourself.
- Image dimensions must be multiples of 16, max edge 3840, aspect ratio within 3:1. Requests outside those bounds are rejected rather than clamped.
- Piping an install script straight to a shell is convenient, not careful. Release binaries are on GitHub if you would rather inspect first.
codex mcp-servermakes Codex controllable by another agent — powerful, and worth understanding before enabling.- Codex moves fast; the CLI is Rust and Apache-2.0, so when docs and behaviour disagree, the repo is the answer.
Tips
codex exec for automationThe non-interactive surface. Run a task, capture stdout, exit with a status code — the right entry point for CI, cron and scripts.
/import from Claude CodeBrings your setup, project config and recent chats across. If you are evaluating both, this removes most of the switching cost.
Write an
AGENTS.md/init generates one. It is a cross-tool open convention, so the same file guides several agents — not effort spent on one vendor.
Set permissions deliberately
/permissions and /setup-default-sandbox decide what Codex may touch. Choose once, properly, rather than approving prompt after prompt.
/ps and /stopBackground terminals keep running. /ps lists them, /stop kills them all — useful after a session that spawned dev servers.
Size the container up front
Code Interpreter defaults to 1 GB. Pass memory_limit="4g" or higher at creation for real datasets — you cannot resize it later.
Resources
openai/codex
The CLI source. Rust, Apache-2.0 — the authoritative answer when docs lag.
github.com
Codex documentation
CLI, IDE, cloud, config, sandbox and security.
developers.openai.com
Image generation guide
Models, sizes, quality and the Responses API image tool.
developers.openai.com
Code Interpreter
How PDFs, ZIPs and spreadsheets actually get produced via the API.
developers.openai.com
Codex Web
The cloud agent surface, included with paid ChatGPT plans.
chatgpt.com
Claude Code
The closest comparison — and
/import moves your setup between them.
www.ty1er.com
OpenAI ChatGPT
The consumer product, and where a paid plan unlocks Codex.
www.ty1er.com