Five AI tools for one small test this week

This week's useful question is simple: can a new AI tool make one handoff disappear without asking you to trust a black box? These five picks cover screen demos, private voice work, coding, model orchestration, and sensitive files. Each one gets a small test, a limitation, and a verdict.

ScreenCursor turns a screen recording into a finished demo while you record. Every click, drag, and keystroke becomes a timed zoom, and you can still retime or delete any zoom before exporting an MP4 at up to 1080p. The official site says the Chrome extension works on macOS, Windows, and Linux, runs offline, and costs $49 as a one-time purchase. 1

The small test is a two-minute product walkthrough. Record it once, change one zoom, and export the result. Try ScreenCursor if your bottleneck is turning a raw screen capture into something people can follow. Wait if you need webcam overlays, phone capture, or a full desktop editor; the site names those as areas where Screen Studio does more. Product Hunt listed ScreenCursor as launching this week. 2

Epilude takes the local-first route for voice. Its Mac app turns speech into formatted text in any text field, and its Notetaker records from your Mac instead of sending a bot into the meeting. The company says meeting notes become an editable transcript, summary, and action list saved as Markdown files you own. Local Mode keeps voice and text on the Mac, even when Wi-Fi is off. 3

Test it on one meeting you already know well. Compare the transcript with the decisions and owners you wrote down yourself, then open the Markdown file outside Epilude. Try it if privacy and editable notes matter more than a large cloud collaboration layer. Wait if you need Windows support or a mature team workspace; the current product is for Mac. Product Hunt listed the Notetaker launch this week, and Epilude’s own update page dates its local EP-1 model to September 8. 45

SWE-2 is the coding pick. Cognition says the model reaches 50 percent on FrontierCode 1.1 Main, within one point of Fable 5.1 at 64 percent lower cost in the company’s comparison. The useful change is not only the score. SWE-2 is available in Devin Desktop and CLI, and Cognition reports that its medium effort level took 58 percent fewer turns than SWE-1.7 on that benchmark. 6

Use a disposable repository and give it one issue with a clear test. Inspect the diff, the tests, and the number of exploratory steps before you let it touch a real codebase. Try SWE-2 if you already work inside Devin and want a cheaper coding-agent run. Wait if you need an open-weight model, a standalone API, or a result independent of Cognition’s benchmark setup. Cognition published the launch on September 10. 6

Sakana AI is making a different bet: let several models cooperate instead of choosing one model for every task. Fugu Max routes a task to the leanest capable model to push down token cost. Fugu Ultra v2 aims at the other end of the curve, where sustained reasoning and difficult software work matter more than the cheapest run. Sakana says both are available through its OpenAI-compatible API, and an existing Fugu integration needs a one-line parameter change to upgrade. 7

The test is a two-part job: ask Fugu Max for a routine research or coding task, then send a harder multi-step version to Ultra v2. Compare the output quality, token bill, and how much verification the answer needs. Try Fugu when model switching and vendor resilience are part of your engineering problem. Wait if you want a simple single-model endpoint with behavior that is easier to attribute. Sakana announced the release on September 11 through its official account. 8

Hybrid Compute on Mac splits an agent task between two places. Perplexity says the cloud handles frontier reasoning, web search, and planning, while a local model handles private files and on-device actions. A privacy gate can keep sensitive details local, mask them, refuse the action, or ask for consent before information reaches the cloud. The feature requires an Apple silicon Mac running macOS 15 or later with at least 24 gigabytes of unified memory, and it is available to Pro, Max, and Enterprise subscribers. 9

Try it with a public research task that also needs one private document. Check what the local model sees, what the cloud model sees, and what the privacy gate records. Try Hybrid Compute if your work needs web research alongside client files or other sensitive material. Wait if you do not have a compatible Mac or if your organization needs a policy and audit surface that it has not configured yet.

The common thread is not that these tools are smarter in the abstract. ScreenCursor removes the editing handoff. Epilude removes the cloud upload and the note-taking pass. SWE-2 reduces some coding-agent detours. Fugu chooses machinery for the task. Hybrid Compute chooses where the sensitive part runs. Pick the one bottleneck you can measure this afternoon, and keep the result where you can inspect it.