Everyone gets stuck mid-task: an error you don’t understand, a config value you can’t find, a form you don’t know how to finish. Usually you stop, stare at the screen, then switch to a search engine or a chat window and explain the whole situation from scratch. That costs time and breaks your focus.
Next-step hints removes that gap. It is an opt-in feature for screenpipe that notices when you stall mid-task and shows a small card in the corner of your screen with the likely next step and a few short how-to lines. The card has two buttons:
I’ll do it collapses the card into a checklist you follow yourself.
Do it for me hands the step to screenpipe’s agent, which carries it out and reports back in one line.
A hint only appears after a real pause that follows active work, for example 20 seconds without input after a minute of typing. It never shows while you’re typing, in a meeting, watching video, or in a blocked app such as a password manager or banking site. It is rate-limited to one hint every few minutes.
Because screenpipe already knows what has been on your screen, you don’t have to explain anything. In our demo it spots a missing STRIPE_KEY from the terminal error and fixes .env.local, then fixes an invalid currency in a report config so the page rebuilds.
Safety is built in. Screen text is treated as data, never as instructions. Nothing runs without a click, and actions like sending, paying, deleting or pushing always need a separate Confirm.
The feature has two halves.
In the screenpipe engine (Rust): a stall detector polls screenpipe’s input activity clock once a second, using lock-free atomic reads that stay off the capture hot path. When it detects a pause after active work, it reads the recent frames from screenpipe’s hot frame cache to confirm you were working in the same app. It then emits a next_step_stall event on screenpipe’s event bus.
In the desktop app (Tauri, Rust + TypeScript):
A gate drops the event if any rule fails: meetings, blocked apps, rate limit, quiet hours.
A context builder pulls the screen text of the focused window, the windows you visited in the last 10 minutes, and your recent clicks from screenpipe’s local API. It never includes keystrokes, and it runs the bundle through screenpipe’s PII filter before anything leaves the device.
One LLM call through screenpipe’s local chat proxy (Ollama, your own API key, or the cloud gateway) returns a strict JSON hint, which is validated before display.
The card is rendered in screenpipe’s notification panel.
Do it for me starts a pi agent session that requires approval before running commands.