ChatGPT Computer History Logs Clicks and Keystrokes on macOS

OpenAI shipped Computer History on August 16, 2026 as an opt-in feature inside the ChatGPT desktop app for macOS, giving the application a persistent record of your clicks and keystrokes that ChatGPT and Codex can query when you make a request. For engineers already thinking about what the agentic era means for data science, this is the moment desktop automation moves from API experiment to consumer surface — with all the privacy architecture questions that entails.

What Computer History captures

Rather than taking screenshots the way Windows Recall does, Computer History records "events" — a deliberate architectural distinction OpenAI is leaning on to differentiate from Microsoft's approach. The feature does not capture images, video, or audio. It builds a timeline from interaction events, and that timeline is the data surface both ChatGPT and Codex can reference when a user asks about prior work. In a demo, Dominik Kundel, a member of the OpenAI Developer Experiences team, showed the feature retrieving the last document he edited, verifying whether it had been shared via Slack, and producing a recap of his morning's activity — all from the event log rather than any stored visual frame.

OpenAI product and engineering manager Ari Weinstein confirmed on X that content inside incognito or private browser tabs is automatically excluded from capture.

User controls

The feature launches opt-in rather than opt-out, a meaningful policy choice given the sensitivity of keystroke-level telemetry. Users can additionally blocklist specific apps and websites from Computer History entirely, and individual entries can be deleted for granular retroactive control. The combination of opt-in enrollment, per-app exclusions, and entry-level deletion gives the privacy model three independent control planes — more layered than the initial Windows Recall rollout, though the comparison only holds so far given the different data modalities involved.

Computer History vs. Windows Recall

The event-stream approach is the core technical differentiator. The table below maps the two features against the dimensions that matter most to developers building on top of either platform.

Developer and security implications

Granting a model access to a structured event log of every app interaction is a qualitatively different capability from granting it access to document content. Codex being able to observe the sequence of actions a developer takes — which files were opened, which terminal commands were run, which Slack threads were checked — creates an automation suggestion engine with far richer context than clipboard or file access alone. That same richness is a new attack surface: a compromised or prompt-injected Codex session could potentially read workflow patterns in ways a sandboxed code-execution environment cannot. The agent hijack risks already documented around coding agents apply with additional force when the agent has a timestamped record of inter-application behaviour.

Computer History signals that OpenAI is building toward persistent, cross-session desktop agency rather than isolated chat turns. The event-stream architecture — no screenshots, structured exclusions, Codex integration — suggests the company is engineering toward regulatory defensibility as much as capability, positioning the feature to survive the scrutiny that forced Microsoft to redesign Recall before its public launch.