Codex 5 New Features

Codex Quietly Rolled Out 5 New Features: Lock-Screen Agent, One-Key Window Capture, and Multi-Day Autonomous Goals

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

Abstract: In July 2026, OpenAI quietly pushed 5 updates to Codex: lock-screen background execution so it keeps working after you close your laptop, Appshots to feed any window's content to Codex with a double-tap of the Command key, /goal mode for setting multi-day autonomous objectives, team-wide plugin sharing, and an advanced annotation mode inside the built-in browser. These aren't scattered fixes — together they signal Codex's transformation from a coding assistant into a general-purpose autonomous agent. The top three features — lock-screen, Appshots, and /goal — can be chained into a powerful workflow: capture a requirement doc on your phone during the commute, set a /goal, and let Codex work overnight while your Mac is locked. When paired with local AI hardware like Kaihe AIBOX for true 24/7 execution even when the computer is fully turned off, Codex becomes a remote-controllable, always-on AI employee that works while you sleep.

In July 2026, OpenAI pushed an update to Codex. No launch event. No livestream. Not even a high-profile tweet. But the five features that dropped represent one of the most significant functional leaps Codex has made since launch: it's no longer a tool you ask questions to — it's an agent that keeps working whether you're at your desk or not.

If you've been treating Codex as just a stronger ChatGPT for coding, this update will change how you see it. We're now looking at an autonomous agent that can run while your screen is locked, understand any on-screen content with a single keyboard shortcut, and set a long-term goal then plan and execute autonomously for days.

This isn't a coding tool update. This is Codex graduating from "AI coding assistant" to "general-purpose autonomous agent."

Feature 1: Lock-Screen Background Execution — Locked Screen ≠ AI Clocked Out

The first feature, and the one most easily underestimated: Codex can now continue running in the background after you lock your Mac screen.

How to use it: Make sure the Codex desktop app is running, lock your Mac normally, then connect via the Codex remote control interface on your phone and send commands.

Sounds like just "background mode," right? Think about what it actually means.

Before this update, Codex was like a coworker who sits next to you. You open your laptop, log in, type your command, wait for it to finish, check the result. Close your laptop to go to a meeting, and Codex clocks out too. Every AI tool had this physical bottleneck — AI stops when the computer stops.

Lock-screen background execution breaks that physical constraint. You can send Codex instructions from your phone during your commute, and it keeps working on your Mac at home. You can think of a task while lying in bed, grab your phone and tell Codex, and wake up to find the results already on your desktop.

This is what AI should feel like — you tell it "do this thing," and you don't stare at a progress bar. You go to a meeting, go to dinner, go to sleep, and when you come back, the work is done.

Pair this with a local AI workstation like Kaihe AIBOX, and the scenario goes even further. Codex's lock-screen mode solves for "AI keeps running when the Mac screen is locked." Kaihe AIBOX solves for "AI keeps running when the computer is completely off." Your Mac, even when locked, is still powered on and consuming electricity. Kaihe AIBOX is a 24/7 standalone hardware device — you shut down your computer, unplug the power cord, and the AI box is still running batch tasks, monitoring logs, and processing scheduled jobs. Lock-screen execution + local AI box = You can go completely offline; AI stays permanently online.

Feature 2: Appshots — Double-Tap Command, AI Instantly Reads Anything On Your Screen

The second feature, Appshots, addresses a long-standing pain point: how do you quickly get the AI to understand what you're looking at?

Before, you'd copy-paste webpage content into Codex, or take a screenshot and upload it, or write a long description. Every approach hit the same input-efficiency wall — you were helping the AI "read" instead of letting the AI "see" for itself.

Appshots changes that. You're browsing a webpage, a Figma design, or an Excel spreadsheet. Double-tap the Command key (customizable in settings), and Codex instantly captures the current app window's complete state — including all text, data, and even content in scrollable areas not currently visible on screen — and sends it directly as conversation context.

No selection. No copy. No screenshot. You just look at what you want the AI to understand, and double-tap Command.

What does this unlock? You're reviewing a GitHub PR — double-tap Command, Codex reads the entire diff and starts a code review. You're looking at a competitor analysis report — double-tap Command, Codex reads it and extracts key insights. You're staring at an error log — double-tap Command, Codex reads it and tells you what went wrong and how to fix it. The friction of information transfer disappears — you no longer think about "how do I feed this to the AI," you just switch to the window and double-tap.

Feature 3: /goal — Set One Objective, Let AI Execute Autonomously for Days

The third feature is the most futuristic: /goal long-term objective mode.

How to use it: In a Codex conversation, begin with /goal and clearly describe your objective, scope, and constraints. For example: /goal Refactor the user authentication module from JWT to OAuth 2.0, covering login, registration, token refresh, and session management, maintaining compatibility with the existing user database.

When Codex receives a /goal, it doesn't start coding immediately. It does three things first: decomposes the objective into executable sub-tasks, plans execution order and dependency relationships, and estimates complexity and time for each sub-task. Then it starts executing autonomously. You don't need to come back after every step to give the next instruction — Codex plans, executes, and verifies on its own. You can type "progress" at any time to check status, or "pause" / "adjust direction" to intervene.

A /goal can run for hours or even days. During that time, you can do other things, lock your Mac screen (thanks to Feature 1), and check progress from your phone. Codex becomes a project executor that you give a "weekly goal" to and review every Friday — not a code typist that needs you to dictate every line.

This is fundamentally different from traditional AI tools. Traditional AI changes one file when you tell it to change one file. Codex's /goal is you giving it a direction, and it plans its own path, executes, and self-corrects. This capability directly maps to a junior engineer's daily workflow — "this module is yours, finish it this week."

Feature 4: Team-Wide Plugin Sharing — Configure Once, Sync to Everyone

The fourth feature targets team collaboration: admins can share custom-developed Codex plugins to all team members with one click.

Before, every team member had to install and configure the same plugins individually. Someone used the wrong version. Someone missed a configuration step. Someone didn't even know the plugin existed. Team toolchain consistency was maintained entirely by word of mouth. Now, an admin sets a configured plugin to "team shared" in the Codex plugin management interface. All team members refresh their plugin list in their client, and it's there. Unified version. Unified configuration. Zero manual setup per person.

This is a massive efficiency boost for enterprise teams using Codex. Especially for teams that have developed custom MCP tools or internal API plugins — instead of writing documentation teaching everyone how to install it, the admin clicks once and it syncs to everyone.

Feature 5: Advanced Annotation Mode — Visually Mark Up Web Pages Inside Codex

The fifth feature targets frontend development and design collaboration: Codex's built-in browser now supports an "annotation mode." Open the Codex built-in browser, navigate to the target web page, and enable annotation mode. Then you can work like you're in Figma — drag webpage elements to adjust layout, and add visual annotations and text comments directly on the page. Before, you'd tell a designer "move that button a bit to the right," and they'd move the wrong button. Now you circle the button on the page, write "move this 20px to the right," and both AI and designer see the same annotated view. Communication cost drops to zero.

These Five Updates Point in One Direction

If you connect these five updates, they aren't scattered features — they form a clear roadmap: lock-screen execution means Codex is no longer a "tool" that clocks in and out with you, but a permanent "colleague"; Appshots means Codex no longer needs manual descriptions, it directly "sees" your work context; /goal means Codex no longer needs step-by-step direction, it receives an objective and plans autonomously; plugin sharing transforms Codex from a personal tool into team infrastructure; annotation mode evolves Codex from pure text interaction to visual collaboration.

Taken together, Codex is transforming from a "you ask, it answers" coding assistant into a "you say what you want, it figures out how to deliver" autonomous agent.

The three core features — lock-screen, Appshots, and /goal — can already be chained into a complete workflow: on your commute home, use Appshots on your phone to capture a requirement doc into Codex, then set a /goal for it to refactor a module. Codex runs all night on your locked Mac. Next morning you open your laptop: code is refactored, tests have passed, a change summary is waiting. You just need to review and merge. Add a Kaihe AIBOX to the mix, and you don't even need to keep your Mac running — the AI box runs independently at less than one-tenth the power consumption of a Mac, 24/7. During the day, Codex helps you write code. At night, the AI box runs CI/CD pipelines, monitors alerts, and generates scheduled reports. You trade your sleeping hours for someone else's workday output.

KaiheAIBOX #AIAgent #OpenAICodex #AICoding #AutonomousAgent

Kaihe AIBOX · Your Private AI Assistant Working 7×24 · AI Agent