typesafe-computer-use and Jev-cu: Two Takes on Computer-Use Agents

Computer-use agents have a shared bottleneck: every click, scroll and keystroke needs a decision about what to do next, and running that decision through a frontier model is slow and expensive. Two projects built in the weeks after Jev's release attack the bottleneck from different ends. typesafe-computer-use reads the whole macOS screen — OCR plus the accessibility tree — and asks Jev to pick from a handful of candidate actions. Jev-cu skips pixels entirely and feeds Jev only the text extracted from a Codex Computer Use session.

Both are among the more heavily-starred entries on the Awesome Jev radar: 861 stars for typesafe-computer-use, 585 for Jev-cu, as of September 2026. Star count is not proof of reliability, but it does mean more eyes have read these READMEs than most of the projects in this series. Here's what those READMEs actually claim, and what they don't.

Key Takeaways

• Different inputs, same shape of decision. Both projects reduce "what happens next" to a small multiple-choice question Jev can answer in milliseconds, instead of a paragraph a bigger model has to write.
• typesafe-computer-use publishes a real cost/latency table — $0.0002 vs $0.032 per decision against Claude Opus, and 0.13–0.38s vs 5.2s per step — but it's a single-setup comparison, not an independent benchmark.
• Jev-cu defaults to dry-run. Deletion, payment, sending, and similar high-risk operations stop for human confirmation unless you turn that off.
• Text-only has real blind spots. typesafe-computer-use admits OCR can't read icons and accessibility coverage varies 0–100% by app; Jev-cu can't see anything a Codex session doesn't already expose as text.
• Neither is production-hardened. typesafe-computer-use flags canvas UIs and bitmapped text as invisible to it; Jev-cu whitelists only specific apps in scripts/policy.mjs.

typesafe-computer-use: OCR, accessibility trees, and a real cost table

typesafe-computer-use (MIT, 861 stars as of September 2026) takes a goal in plain English — its README example is "go to techcrunch and take me to the checkout page for the cheapest tickets" — and drives the whole macOS desktop toward it. The pipeline is: screen capture → Vision OCR → block merging → accessibility tree walk → item list. To keep this fast, it crops to the active window plus menu bar and reuses prior OCR results by comparing tiles; if fewer than 60% of tiles changed since the last step, it skips re-reading the unchanged ones.

At each step Jev is handed 2–3 mutually exclusive choices — click this, type that, scroll — and returns a full probability distribution with a calibrated confidence, in the author's words "a few hundred milliseconds." The README's headline table compares this to Claude Opus doing the same decision: $0.0002 vs $0.032 per decision (about 155× cheaper), and 0.13–0.38s vs 5.2s per step (14–40× faster). A browser-specific backend that reads the DOM instead of OCR reports 302–380ms end to end. These are the author's own numbers from their own setup — not independently retested here, and not a guarantee your app's UI will OCR as cleanly as TechCrunch's checkout flow.

The limitations section is candid: OCR can't read icons, accessibility tree coverage ranges from 0% to 100% depending on the app, only the primary monitor is captured, canvas-based UIs and bitmapped text are invisible to both OCR and the accessibility tree, and passwords are never typed by the model — it relies on your password manager or SSO buttons instead. Aiskill.market entry: /skills/typesafe-computer-use-awlevin.

Jev-cu: text-only, dry-run by default

Jev-cu (MIT, 585 stars as of September 2026) takes a different path. Instead of screenshots, it runs inside a Codex Computer Use loop and passes Jev only the text extracted from interface elements — labels, button text, form fields. Jev decides which element to interact with, what action to take, whether the goal looks complete, and whether the step looks risky; Codex still does the actual reading and executing. Splitting the loop this way means Jev is answering "what's next" while Codex handles "what does the screen look like" and "how do I click it."

The safety default is the most notable design choice: dry-run is on by default. Any operation involving deletion, sending, payment, authorization, file upload, verification codes, installation, or system settings changes pauses for a human confirmation step before it executes. The project also whitelists applications explicitly in scripts/policy.mjs, rather than assuming it's safe to act on anything Codex can see. The README is explicit that Jev-cu does not bypass logins, paywalls, or verification codes, and treats interface text strictly as data — not as instructions to follow, which matters if a page contains adversarial text. Aiskill.market entry: /skills/jev-cu-sac-y.

Reading the two side by side

The practical difference is what each one needs to already be running. typesafe-computer-use wants a Mac and nothing else — it reads the screen directly. Jev-cu wants an existing Codex Computer Use session and rides inside it, which is a smaller lift if you're already a Codex user but a hard dependency if you're not. Neither project publishes a task-success rate against a shared benchmark, and neither has been independently re-run for this article — both are honest about that in their own docs, which puts them ahead of a lot of the field. For the browser-only equivalent of this same idea, see our earlier writeup of Jev-Ultrafast; for the wider browser-agent family that swaps OS-level access for a Playwright session, see more Jev browser agents.

What to do next

1. Match the tool to what you already have running. If you live in Codex Computer Use, Jev-cu adds decisions to an existing loop. If you want a standalone Mac automation, typesafe-computer-use is the more complete pipeline.
2. Treat the cost/latency tables as directional, not guaranteed. Run your own comparison on your own UI before trusting a 150×+ savings claim for your workload — see how to evaluate Jev projects honestly.
3. Keep dry-run and confirmation gates on for anything touching payment, deletion or account settings, regardless of which tool you pick.
4. Check accessibility coverage on your target app first. typesafe-computer-use's own README warns coverage swings from 0% to 100% — test on the actual app, not a demo.
5. Watch the radar for convergence. With two teams solving the same problem two different ways within weeks, a third hybrid approach (screenshots plus a Computer Use connection) seems likely; Awesome Jev is the place to see it land first.