Codex (OpenAI)

OpenAI's coding and agentic work surface, available through the Codex app, VS Code extension, terminal, and a Claude Code plugin. In Nate Herk's coverage, Codex has evolved from a coding agent into a ChatGPT-like project workspace that can read local files, create spreadsheets, automate browsers, use plugins, build skills, deploy dashboards, and run scheduled automations.

Key facts

• Powered by GPT-5.4
• Beats Opus 4.6 on most coding benchmarks (13-point lead on some)
• Opus 4.6 slightly leads on SWE-bench verified
• Free with a ChatGPT subscription
• Known weaknesses: less strong on planning, more rigid, fewer creative outputs
• Known strengths: catches bugs Opus writes, better at code review, stronger UI output in one-shot prompts
• Nate's May 2026 Codex course positions Codex as a "super app" for local project work: chats, projects, model/reasoning controls, file access, browser automation, plugins, skills, scheduled automations, and deployment workflows ^[src-048]
• Codex can work inside an existing Claude Code-style AIOS directory if the project has clear instructions, folders, skills, agents, and settings, supporting Agent Harness Portability ^[src-048]
• Nate's working preference: use Claude Code for exploratory planning and creative ideation; use Codex for pragmatic execution, long-plan following, troubleshooting, and review ^[src-048]
• The tutorial demo uses Codex to build a YouTube comment intelligence system: pull comments, analyze patterns, write an Excel workbook, create visualizations, build a dashboard, and schedule weekly refreshes ^[src-048]
• Codex exposes slash commands and plugin/tool surfaces including autoreview, code review, feedback, MCP, memories, model, plan mode, reasoning, browser use, GitHub, Figma, HyperFrames, Canva, Google Drive, Slack, SharePoint, Teams, and others ^[src-048]
• Nate's May 8 stack video places Codex in A tier: a weekly companion rather than his primary daily driver, useful because its strengths complement Claude Code's weaknesses and vice versa ^[src-053]
• Jack Roberts's Codex walkthrough emphasizes five interfaces, project folders, plugin connectors, skill creation, automations, Chronicle screen memory, and a Multi-Brain Model Strategy where Codex coordinates with Lovable, Claude, and Gemini ^[src-058].
• OpenAI's Tibo Sio says Codex has moved beyond coding: the majority of tasks being performed in Codex are non-coding tasks, including document search, data questions, launch coordination, daily briefs, file manipulation, spreadsheets, slides, websites, and background automations ^[src-081].
• Sio describes Codex's slash-goal mode as a long-horizon mode where the agent can pursue a hard objective for hours, days, or weeks until it can judge the goal satisfied ^[src-081].
• Enterprise trust features include sandboxed file access, optional network restrictions, read-only scopes, and auto-review agents that can stop risky actions by the primary agent ^[src-081].
• OpenAI's February-May 2026 videos show Codex used by PMs and engineers for understanding code, making changes, fixing CI, querying logs, updating skills, running tests, working in worktrees, and validating its own output before handoff ^[src-084].
• The API & Codex Build Hour frames Codex as part of a broader Harness Engineering shift: agents become reliable when repositories include clear task entry points, validation harnesses, docs, skills, standards, and decision records ^[src-084].
• The May 2026 Chrome extension video adds a browser-work surface: Codex can use the user's real Chrome profile, tabs, cookies, plugins, code execution, and parallel tab groups when structured connectors are not enough ^[src-084].

What it does

In ^[src-048], Codex is framed as a bridge between ChatGPT and a local agent runtime. It keeps the familiar chat/projects interface but adds access to local files, spreadsheets, browser control, reusable skills, plugins, deployment targets, and recurring automations.

The course's practical pattern is to start in plan mode, let Codex inspect the project, align on a written plan, set up any required credentials or APIs, then allow implementation. Permissions matter: Nate recommends starting conservatively, then moving to auto-review or full access only once the user understands what Codex is allowed to change ^[src-048].

The most strategically important claim is portability. Nate uses the same local project directory across Codex and Claude Code: if the directory contains clear conventions and instructions, the harness can be swapped while preserving the work system ^[src-048].

In ^[src-053], Codex becomes part of a broader Lean AI Tool Stack rather than the center of the system. Nate still does most work in Claude Code, but reaches for Codex when its execution, review, or troubleshooting strengths fit the task ^[src-053].

In ^[src-058], Roberts pushes the "super app" framing further. Codex becomes a persistent workspace with folders for life/business domains, pinned threads, parallel agents, skill creation, plugin connectors, browser/computer use, image generation inside apps, scheduled automations, reminders, and Chronicle-style screen context ^[src-058].

The source also positions Codex as the central workbench in a multi-tool build loop: generate a polished UI in Lovable, sync to GitHub, clone and enhance it in Codex, and invoke Claude or Gemini through terminal/plugin access when those tools are better suited to design or long-context work ^[src-058].

OpenAI's own Forum discussion in ^[src-081] reinforces the same direction from inside the Codex team. Sio frames Codex less as a code generator than as a general computer-work agent: it can gather information, manipulate files, create artifacts, run recurring checks, ask people for status, and use code as an internal tool even when the user never reads or writes code.

OpenAI's own product demos in ^[src-084] make that shift more concrete. Codex is shown as a project manager's code-understanding tool, an engineer's validation loop, a CI/log investigator, a skill-improvement surface, a worktree manager, and a Chrome-connected browser worker.

Source references

• ^[src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)

– Videos referenced: B2Kh_ZoLVTM

• ^[src-013] Nate Herk — "Build & Sell Claude Code Operating Systems (2+ Hour Course)" (2026-05-01)

– Portability test: Nate migrated his full AIOS to Codex in 2 minutes to validate that Claude Code skills are harness-portable. "I took my AIOS and I moved it over to Codex just to make sure that everything was good, and it took Codex about 2 minutes to get adjusted."

• ^[src-048] Nate Herk — "Master 97% of Codex in 1 Hour (full course)" (2026-05-06)
• ^[src-053] Nate Herk — "Overwhelmed By AI? Just Copy My Tech Stack" (2026-05-08)
• ^[src-058] Jack Roberts — "How to use Codex Better than 99% of People" (2026-05-06)
• ^[src-081] OpenAI — "Codex for Everyday Work: AI Agents Beyond Coding" (2026-05-14)
• ^[src-084] OpenAI Codex, Workspace Agents, Prompt Caching, and Superintelligence Policy cluster (2026-02-09 to 2026-05-08)