ChatGPT's Computer History Tracks Your Clicks and Keystrokes

OpenAI shipped Computer History for ChatGPT's macOS desktop app on August 14, 2026 — a feature that builds a searchable activity timeline from your clicks, keystrokes, and app switches, giving ChatGPT and Codex a live window into what you've actually been doing all day. It's opt-in, it skips screenshots entirely, and it auto-ignores anything that happens in a private browser tab.

Background You Need

The feature didn't arrive without precedent inside OpenAI's own product line. Earlier, the company had run a research preview called "Chronicle" that worked by capturing periodic screenshots of your screen — essentially a photo album your AI could rifle through when you asked a question. Chronicle had an obvious structural problem: screenshots capture everything indiscriminately, including passwords typed into visible fields, sensitive documents left open, and private messages in the background. The approach was technically legible but privacy-hostile by design, and it never made it to general release.

Microsoft's parallel experiment with Windows Recall created the clearest template for what not to do. When Recall launched in 2025, it was opt-out by default, meaning it activated automatically unless users explicitly turned it off. The backlash was fast and public enough that Microsoft reversed course before the feature reached wide release, converting it to opt-in. The episode established a rough industry consensus: ambient activity tracking needs to default to off, and the data collection needs to be legible to the user rather than buried in settings. Computer History is OpenAI's answer to both lessons simultaneously — it replaces Chronicle's screenshot architecture with an event-based alternative, and it launches opt-in from the first day rather than requiring users to act defensively.

The broader context matters for developers specifically. The question of how AI products should handle user-behavior data has been unsettled since AI assistants moved from cloud-only to desktop-resident. Computer History is the most prominent attempt to date to define what a privacy-respecting activity-tracking model actually looks like in practice.

What's New

Computer History reads your activity through macOS's accessibility system — the same OS layer that screen readers and assistive technologies use to observe what's happening on screen. According to The Decoder's Matthias Bastian, who covered the feature's August 14 launch in technical detail, this means the system captures clicks, keystrokes, keyboard shortcuts, and app switches as discrete events. Screenshots, screen recordings, microphone input, and system audio are explicitly excluded. Activity in private or incognito browser tabs is automatically ignored regardless of any user setting.

Those raw events don't persist as-is. The Decoder reports they're periodically converted into text summaries and stored locally as Markdown files — a detail the primary Verge coverage gestures at but doesn't fully specify. The timeline interface groups those summaries by day and time, showing which applications were active during each window. The result is closer to a structured behavioral journal than a surveillance log: readable by the user, queryable by the AI, and built from signals rather than images.

Multiple reports indicate the feature has a workflow-automation angle that goes beyond simple recall. When ChatGPT detects repeating patterns in the recorded activity — the same sequence of apps opened each morning, a consistent document-review loop, a regular file-move ritual — the timeline surfaces those patterns as candidate "Skills" that Codex can convert into reusable automations. VocoLife's summary of the feature notes that a demo showed the app retrieving a recently edited document, checking its sharing status, and generating a morning activity recap — exactly the kind of lookup that currently requires manual search or memory.

User controls are explicit rather than buried. Specific applications and websites can be excluded from tracking entirely, individual timeline entries can be deleted after the fact, and private-tab exclusion is automatic. An OpenAI manager confirmed the private-browsing behavior directly, per multiple reports. These controls collectively position Computer History as a user-controlled personal log rather than a background telemetry system — a meaningful design choice, even if its sufficiency is debatable.

The Pushback

Our read is that the privacy gap the launch coverage leaves open is the one that matters most: what happens when you actually query the timeline. The Decoder confirms the Markdown summaries are stored locally, but "recap my morning" is a server-side call by definition — and none of the sources covering this launch clarify whether the context pulled from your local event log travels to OpenAI's servers as part of that request. If it does, the "stored locally" framing is accurate but incomplete. The accessibility-layer architecture also raises a question about sensitive inputs. macOS's accessibility system has a privileged view of everything you type, including passwords and form fields. How Computer History filters — or doesn't filter — those inputs wasn't addressed in any of the launch materials across the four sources covering this story, and that's the kind of omission that tends to get answered by security researchers rather than by press releases.

The comparison to Windows Recall that multiple reports invoke is also worth stress-testing rather than simply accepting. Recall's controversy wasn't only about its opt-out default. Security researchers demonstrated after launch that Recall's local database of screenshots and OCR text could be exfiltrated by malware operating with standard user permissions. Computer History eliminates the screenshot attack surface entirely, which is a genuine improvement, but a locally stored log of behavioral events dense enough to reconstruct daily workflows is still a target worth thinking about. Finally, there's a platform gap the competitive framing obscures: Computer History is macOS-only at launch. No Windows version has been announced, despite the fact that Recall is a Windows-first feature and the entire context of this launch is an implicit response to Recall's design decisions. The accessibility-layer approach works on macOS because of Apple-specific APIs; replicating it on Windows would require a different architecture. Whether OpenAI has that path is currently unanswered, and it's a meaningful gap if the goal is to offer a cross-platform alternative to what Microsoft built.

Sources