Record it once.
Let AI replay it — forever.
Taskloopr watches how you work in the browser, then replays it with self-healing element matching, an AI that clicks what you meant, and plain-language edits — no code, no brittle CSS selectors, no "it broke because the button moved."
Fallback depth
8-tier
self-healing chain
Trusted input
Real OS-level clicks via CDP
3 results
TasklooprWorkflow finished
now
6/6 steps completed — self-healing kicked in once.
Needs your approval
2m ago
AI Decision: Proceed — 92% confidence. Resume the run?
Click to see it record →
Live step feed
Everything you'd want an agent to do on your behalf
Four real AI capabilities layered on top of a deterministic recorder — each one logged, each one honest about when it fell back.
Record & replay, exactly as it happened
Clicks, typed values, Enter-to-search, and full-page or SPA navigations — captured as you browse, then replayed at a human pace.
AI Decision steps
A real OpenAI-backed proceed/skip check, multimodal — it reads page text and a live screenshot before deciding.
AI-generated values
Toggle it on, write an instruction, and every run types a fresh value — logged as (AI-generated), never silent.
AI-driven click
"Click the first real result, skip anything sponsored." A vision model looks at the live page and picks the right element.
Conversational editing
"Remove the like-button step." Describe a change in plain language — it edits the workflow directly, no manual toggling.
Trusted input (CDP)
Replay dispatches real, OS-level input events via the Chrome DevTools Protocol — indistinguishable from a real person.
Run history & alerts
Every run is a real, persisted log — which fallback self-healed a step, and a genuine desktop notification when it finishes.
Dataset loops
Bind {{row.column}} into any step and loop the whole workflow once per row — pasted, CSV, or a manual list — with results exported when it's done.
Runs while minimized
Replay keeps going with the window minimized or the tab unfocused, and survives the service worker restarting mid-run.
A page redesign shouldn't break your workflow
A moved button, a regenerated id, a new class name — ordinary automation stops dead. Taskloopr walks a real fallback chain, eight strategies deep, until one of them finds your element with confidence.
Every adaptation is visible. The run log flags exactly which tier matched, so a silent drift is never invisible.
"Position only" mode. When a search result genuinely varies, id/text/aria-label all describe the old item — position-only matching skips them on purpose.
Optional steps. Mark an incidental click "skip if it fails" so one flaky step never stops the whole run.
Fallback chain
matching…id and data-testid changed after a deploy — replay adapted automatically and the step succeeded.
From click to automation in four steps
No scripting. No DOM inspector. Just do the task once.
Record
Open any real website and tap Record. Browse normally — click, type, press Enter, follow a link. Each becomes a step live, including page navigations.
Save
Finish recording, name the workflow, save. No selectors to write, no config file — just what you actually did.
Run
Open it and tap Run. Taskloopr jumps back to the start and replays every step, outlining each element in blue as it goes.
Self-heal & adapt
The page changed since you recorded? The fallback chain finds it anyway — check the Runs tab for exactly what adapted.
Four ways the AI actually earns its keep
Not a chatbot bolted on top — each capability is a specific, logged operation on top of your recording.
Compare this job description with my resume. Only proceed if at least 70% of the required skills match and the location is India.
AI decision: Proceed (92% confidence)
"React, TypeScript, and 4 of 5 required skills are present. Location listed as Bengaluru, India — matches."
No vaporware. Here's exactly what's real.
This is a dev build, and we'd rather tell you what doesn't work yet than let you find out mid-workflow.
Real today
- Recording real clicks, typed values, Enter-to-search, and full-page/SPA navigations
- Saving and replaying workflows with the 8-tier self-healing fallback chain
- Replay survives a minimized window, an unfocused tab, and the service worker restarting mid-run
- AI Decision, AI-generated values, and AI-driven click — all OpenAI-backed and logged
- Dataset loops with {{row.column}} binding, extract steps, and CSV export
- Pause-and-resume mid-run for a login screen or CAPTCHA
- PDF and plain-text parsing for AI Knowledge
On the roadmap
- Word/.docx files for AI Knowledge (PDF and plain-text formats are read today)
- OCR for scanned PDFs with no text layer
- Freeform natural-language rewrites beyond the fixed edit operations
Get Taskloopr running in under two minutes
No account, no server, nothing to configure to get started — the recorder works the moment it's loaded.
Request the build
Ask for the current dev build and unzip it anywhere on disk.
Enable developer mode
Open chrome://extensions and flip on Developer mode, top right.
Load unpacked
Click "Load unpacked," select the folder, then pin the icon to your toolbar.
Manifest V3 · ~180KB · works offline once loaded
Questions, answered honestly
Do I need to write selectors or any code?
No. You record a workflow by browsing normally — clicking, typing, following links. Taskloopr captures the steps and figures out how to re-find each element on its own.
What happens when the page layout changes after I recorded?
Replay walks an 8-tier fallback chain — id → data-testid → saved position → aria-label → exact text → similar text → toggle label → the only visible match on the page — until one of them resolves. Every adaptation is flagged in the run log, never silent.
Where does my OpenAI key live?
In this browser's local extension storage only. It's never written into a file and never sent anywhere but api.openai.com — required for AI Decision steps, AI-generated values, and AI-driven clicks.
Does it work on any website?
Yes, on any normal website — Chrome blocks extensions on internal chrome:// pages by design, so those are the one exception.
What is "trusted input," and why does Chrome show a banner while it runs?
Replay can dispatch real, OS-level input via the Chrome DevTools Protocol so pages can't tell it apart from a person (isTrusted: true). Chrome shows an owned "is debugging this browser" banner for the duration — that's Chrome's own security behavior for any extension using chrome.debugger, and there's no way to suppress it by design.
Does it keep running if I minimize the window or switch tabs?
Yes. Replay keeps going while the window is minimized and while the tab is unfocused, and it survives the Manifest V3 service worker being torn down mid-run. A 50-step flow replaying in a minimized window is part of the test suite — the page counts its own clicks and all 50 land, in order.
Can I run one workflow over a list of inputs?
Yes. Point a workflow at a dataset and reference columns as {{row.column}} anywhere a value is typed, and it runs once per row. You can also add extract steps that pull values off the page during the run, then export everything as CSV at the end.
What happens if a run hits a login screen or a CAPTCHA?
Add a pause step. The run stops, tells you it is waiting, and picks up where it left off once you have signed in or solved the challenge. AI checks can also be set to alert you and wait rather than guessing or aborting.
What can't it do yet?
Word/.docx files for AI Knowledge — PDF text and plain-text formats are read, but scanned PDFs with no text layer are reported as unreadable rather than OCR'd. Freeform natural-language rewrites are also still limited to a fixed set of edit operations. See "Built in the open" above for the full, current list.