OpenAI’s Codex app for Mac just got eyes on your desktop. Chronicle, its new research preview, snaps periodic screenshots, ships them to the cloud for processing, and turns them into text summaries that stick around locally. Developers get context without endless re-explaining. But those summaries? Unencrypted Markdown files anyone can read. The Next Web broke the details first, noting how background agents capture display content on Apple Silicon Macs running macOS 14 or later.
Only ChatGPT Pro subscribers at $100 a month or more can touch it. No dice in the EU, UK, or Switzerland—GDPR stands in the way. OpenAI president Greg Brockman called it “an experimental feature giving Codex the ability to see and have recent memory over what you see, automatically giving it full context on what you’re doing. Feels surprisingly magical to use.” Raw screenshots hit a system temp folder, vanish after six hours. Servers process via OCR and visual analysis, then delete. Memories persist in ~/.codex/memories_extensions/chronicle/. OpenAI’s docs confirm: no server storage post-processing, no training data use.
And the risks. Prompt injection jumps because context pulls from screen summaries. Sensitive info lingers in plain text. Rate limits burn fast for Pro users. Pause it before meetings—tap the menu bar icon. Needs Screen Recording and Accessibility permissions. 9to5Mac explains users can inspect and edit those memories, building on Codex’s core memory from chats.
Codex itself evolved fast. Launched as a coding agent, the April 16 update dubbed it “Codex for (almost) everything.” Now it runs apps with its own cursor—sees screens, clicks, types—in parallel agents that don’t mess with your work. Multiple terminals. SSH to dev boxes. Over 90 plugins. Image generation. In-app browser. Usage doubled post GPT-5.2-Codex in December; one million developers hooked. OpenAI’s announcement pitches it for frontend iteration, app testing, API-less workflows.
But Chronicle fits a pattern. Screen-aware AI exploded, then imploded on trust. Rewind AI, rebranded Limitless, got bought by Meta in December 2025—Mac app killed, captures disabled. Microsoft Copilot shed 39% subscribers in six months over privacy woes; Recall’s encrypted database proved hackable. Screenpipe goes local-first, $400 lifetime, no cloud. Perplexity’s Personal Computer agents a Mac mini with file access, still clouds core smarts. Chronicle picks speed over lockdown, banking on OpenAI’s no-store pledge.
Privacy hawks balk. Unencrypted locals scream risk—family glances your code? Competitor spies files? Cloud transit opens interception doors, even if temporary. OpenAI warns of it all. Yet developers crave the ambient boost. No more “fix the thing in that tab.” Just “fix this.” Gartner sees over 40% of big firms piloting such intelligence by year-end. Apple tests AI glasses. Slack overhauls with agents. OpenAI brews screenless gear with Jony Ive.
So where’s the line? Chronicle demands opt-in trust in OpenAI’s architecture. Pause for secrets. Edit memories. But always-on eyes shift computing norms. Codex agents already roam your Mac background. Chronicle feeds them your recent world. Magical, yes. Creepy? Depends on the user. Pro subscribers test it now; wider rollout looms if feedback holds. Windows? Coming, but Mac leads with accessibility tree hacks—props to Apple’s API for screen readers, now AI fodder. MacStories praises the parallel cursors, no foreground steals.
Industry insiders watch adoption. Three million weekly Codex users signal scale. But Copilot’s drop-off warns: botch privacy, lose the crowd. Chronicle bets utility wins. Local alternatives like Screenpipe gain if cloud falters. OpenAI pushes the superapp vision—chat, code, agents, memory in one. Chronicle cements it. Pause before that board call. Your screen’s now fodder for the machine.
WebProNews is an iEntry Publication