ChatGPT Work Tool and Skill Reference – Comprehensive Overview

What the Codex Tool Reference is and why it matters

The Codex Tool Reference is a publicly hosted inventory that lists 232 callable tool interfaces and 44 complete skill definitions for ChatGPT Work (also known as Codex). It serves as the single source of truth for developers and AI agents to discover, understand, and invoke the exact APIs that power ChatGPT’s extended capabilities—ranging from browser automation to data analytics, GitHub management, and Gmail operations. By exposing TypeScript declarations, tool descriptions, and the full SKILL.md source for each skill, the reference makes the otherwise opaque “plugin” ecosystem transparent, reproducible, and auditable.

Self‑contained overview of the major tool categories

1. Core callable interfaces (223 tools + 9 session controls)

File and process manipulation – apply_patch, exec_command, view_image, write_stdin, update_plan.
Collaboration primitives – collaboration.spawn_agent, collaboration.send_message, collaboration.list_agents, collaboration.interrupt_agent, collaboration.wait_agent.
Skill management – skills__list, skills__read for discovering and loading reusable instruction packages.
Plugin handling – request_plugin_install suggests a plugin when a required capability is missing.
MCP resources – list_mcp_resource_templates, list_mcp_resources, read_mcp_resource expose server‑side data to the model.

2. High‑level functional groups (skills)

Category	Representative skills	Typical use‑case
Documents & visuals	`answers-charts`, `answers-images`, `documents`, `imagegen`, `pdf`, `Presentations`, `Spreadsheets`, `visualize`, `writing-blocks`	Generate charts, edit Word/Google Docs, create raster images, render PDFs, build slide decks, produce spreadsheets, or render interactive visualizations.
Browser & Sites	`control-browser`, `sites:sites-building`, `sites:sites-hosting`, `sites:sites-preview-troubleshooting`	Automate authenticated web interactions with Playwright, build and host static sites, or troubleshoot preview failures.
Context & interaction	`demos:answers-ask-user-input`, `openai-library:library`, `personal-context`, `resolve-recipients`	Prompt the user for missing data, access the ChatGPT Library, maintain conversation continuity, or disambiguate recipient identities before sending messages.
OpenAI, skills & plugins	`openai-docs`, `plugin-creator`, `plugin-management:plugin-management`, `skill-creator`, `skill-installer`, `template-creator`	Retrieve OpenAI product documentation, scaffold new plugins, manage installed plugins, create or install new skills, and generate reusable templates from existing artifacts.
Data analytics	`data-analytics:build-dashboard`, `data-analytics:build-report`, `data-analytics:visualize-data`, `data-analytics:metric-diagnostics`, `data-analytics:market-sizing`	Build KPI dashboards, produce executive reports, validate data quality, diagnose metric anomalies, and estimate market size with transparent assumptions.
Work pets & demos	`work-pets:create-pet`, `work-pets:update-pet`, `work-pets:pets`, `demos:onboarding-setup-pet`	Create, preview, and manage animated ChatGPT Work “pets” that personalize the user experience.
Automations	`automations.create`, `automations.update`, `automations.list`, `automations.peek`	Schedule recurring reminders, condition‑watch tasks, or webhook‑driven automations for Gmail, Slack, GitHub, and other connectors.
GitHub	Over 80 tools covering repository CRUD, issue/PR lifecycle, workflow artifacts, and code search (e.g., `github_create_issue`, `github_fetch_pr`, `github_merge_pull_request`).	Enable end‑to‑end software development workflows directly from the language model.
Gmail	21 tools for label management, message reading, drafting, sending, forwarding, and bulk operations (e.g., `gmail_apply_labels_to_emails`, `gmail_create_draft`, `gmail_bulk_label_matching_emails`).	Allow AI agents to act as personal assistants that organize, compose, and act on email.

How the reference is structured for AI consumption

Skill pages reproduce the verbatim SKILL.md source, ensuring the model can read the exact instruction set without ambiguity.
Tool declarations are provided as TypeScript signatures, which the model can parse to generate correctly‑typed JSON payloads.
Availability flags (session configuration, permissions, installed plugins) are explicitly documented, so the model can gracefully handle missing capabilities.

Community insights from Hacker News comments

@simonw – “See this comment for background (and creation prompt): https://news.ycombinator.com/item?id=49504625#49505116”
The author points to the original creation prompt, confirming that the reference was generated programmatically from the underlying Codex configuration. This provenance reassures readers that the inventory reflects the live system rather than a static snapshot.

@simonw – “The most interesting skill is control-browser. It tells ChatGPT Work how to launch a Playwright instance via its Node.js REPL, then runs nodeRepl.write(await browser.documentation()); to retrieve usage instructions.”
The comment highlights a design pattern where a skill delegates the heavy‑lifting of documentation to a runtime call instead of embedding a massive markdown blob. This keeps the skill concise, reduces duplication, and allows the documentation to stay in sync with the underlying library.

@satvikpendem – “I don’t get how this is different from Codex if Codex can do all the same things.”
The distinction is that Codex refers to the underlying execution engine, while ChatGPT Work (the UI layer) enforces a strict tool‑call contract, provides a curated UI for skill selection, and adds safety checks (e.g., permission gating, token budgeting). The reference makes those contracts explicit.

@montroser – “Can we get one for how to test websites on a regular sized screen? The left sidebar needs to scroll on its own unless you’re on a 4K monitor.”
This request underscores a practical limitation: many browser‑automation skills assume a default viewport. Extending control-browser with a set_viewport parameter would address such UI‑specific testing needs.

@enraged_camel – “Why do AI‑generated websites always look the same? Feels like Bootstrap.”
The comment hints at a broader design bias: many visual‑generation skills (e.g., imagegen, visualize) rely on default style libraries. Introducing diversity parameters (theme, layout grid, design system) into those skills could mitigate the “Bootstrap‑effect.”

@felixgallo – “Using ChatGPT Work given Sam Altman’s history, OpenAI safety changes, and the HuggingFace debacle is a serious mistake.”
While political, this comment reminds developers to evaluate the trustworthiness of the platform and its governance. The open‑source nature of the reference mitigates some risk by allowing independent audits of the tool definitions.

@darepublic – “Some of these work tools can slow things down and waste plentiful tokens.”
Indeed, invoking tools incurs token overhead for request/response payloads. The reference’s explicit max_output_tokens fields (e.g., in exec_command) help developers set realistic limits and avoid runaway costs.

Practical takeaways for developers

Use the reference as a contract – When building a custom ChatGPT Work integration, import the TypeScript declarations directly to guarantee payload compatibility.
Leverage skill reuse – Instead of re‑implementing common patterns (e.g., “ask the user for missing context”), call the existing demos:answers-ask-user-input skill to keep prompts consistent.
Audit permissions – Each tool lists required permissions (e.g., sandbox_permissions, require_escalated). Validate those against your deployment’s security policy before enabling a skill.
Extend with minimal friction – To add a new capability, create a skill-creator skill, scaffold the plugin directory, and register it via plugin-management. The reference already includes a plugin-creator skill that automates this workflow.
Monitor token budgets – Tools like exec_command and functions.wait expose max_output_tokens and yield_time_ms. Adjust these values based on the expected size of command output to prevent unnecessary token consumption.
Plan for UI constraints – If your workflow depends on visual layout (e.g., testing responsive designs), consider adding viewport‑control parameters to the control-browser skill or wrapping it in a higher‑level skill that sets the desired dimensions before navigation.

Conclusion

The Codex Tool Reference is a comprehensive, machine‑readable catalogue of everything ChatGPT Work can do, from low‑level OS commands to high‑level business analytics. By exposing both the tool signatures and the full skill source, it enables developers to programmatically discover capabilities, audit security boundaries, and build robust AI‑driven workflows. Community comments on Hacker News surface real‑world concerns—documentation design, UI ergonomics, token efficiency, and platform trust—that can guide future enhancements to the toolset.