Claude + Jev for Computer Use: Build with Opus, Automate with Jev

🎯 The Unlock: Faster-Than-Human Computer Use 0:00

The framing is "a pretty historic time in AI": combine a frontier model like Claude Opus 5.5 with a System 1 model like Jev, and you get computer use that is faster than a human, at an insanely cheap price — "trainable, cheaper, and more accurate than ever before." The opening demo is Jev wired to Snake, using its probabilities and keyboard tool access to play "really, really well at max speed."

The claim that sets up the whole video: Jev + Claude + browser-use makes computer agents faster than us, and unlocks real-time use cases that were "literally never possible before."

⚡ Why Jev Fixes Computer Use 1:48

The original problem with computer-use agents was speed and cost: you had to wait for the agent to think about which tool to use, make its decision, and pay for a detailed answer explaining why it clicked the button. But computer use is all tool calls — "do this step, then this step, then that step" — it doesn't need real logic or essay-length reasoning.

🧰 The Tool Stack 3:08

The four pieces, all free and open source:

The division of labor: browser-use for building workflows, Playwright for running them once automated. And the goal of an automated workflow, once it's settled, is to not use AI at all — just the tools it has. Claude Opus 5.5 sits back as a fallback: "one of the best in the world at computer use automation, it's just expensive — so we don't want to use it at all times."

🔄 The Method: Build with Claude, Automate with Jev 5:05

The two-phase method is the heart of the video:

• BUILD (with Claude Opus): think about the actual steps needed for the task. Example — a competitor-research/scraping tool: tell Claude your competitors, it uses Chrome CDP to open Instagram, YouTube, TikTok, Twitter and find their recent posts, works out how to automate the flow, and by the end turns it into a SKILL.md file.
• AUTOMATE (with Jev): hand that SKILL.md to Jev as one of its choice options. Now an automation like "send 20 LinkedIn connection requests" makes Jev's probability select that tool and run the whole workflow — at essentially no cost. Claude stays wired in as the fallback API, called only when the UI changes or something goes wrong.

♟️ Example: Jev Beats Chess.com 6:41

The first completed workflow runs entirely on Jev: "on chess.com, start a game and play until you win." The clever bit is the architecture — Jev is connected to Stockfish, which feeds it the best move probabilities based on the board, and because chess is 2D the board state is written back as JSON for Jev to decide from. The decision tree in the corner shows whether it's using Jev or falling back to Claude/Haiku; when it stumbled early it fell back to Haiku to get the game going. ~400 seconds in, it won (the opponent abandoned, but it was already "dominating the board").

🚗 Example: Facebook Marketplace Car Flipper 8:30

The second example is a work-in-progress that shows the full loop: a Facebook Marketplace car flipper that must search "Mercedes-Benz CLA," set a 500km radius around Toronto, cap year at 2020 and price at $15,000, then find a matching car and message the seller. The goal is for Claude Opus 5.5 to learn every step, then filter ~95% of those steps off to Jev to save costs at the same quality.

The demo is genuinely close: it opens Facebook, searches the right model, sets the price and year filters perfectly ("basically a perfect search"), scrolls, and finds a qualifying Mercedes — but the final "message the seller" step came out wrong. The point, shown live rather than hidden: that's exactly when you iterate with Claude in the chat, perfect the workflow, then hand it to Jev and set the automation.

🎯 The Pattern & What Else It Runs 10:30

The reusable pattern, distilled: a frontier model learns the workflow once and writes it as a skill; a System 1 model executes it as a cheap, fast, schema-constrained decision; the frontier model stays on call as a fallback for the cases the routine can't handle. "If you can get Jev to do this, you can get it to almost do anything."

The workflows already running on this channel make the scope concrete: LinkedIn connections at scale, Instagram DMs at scale, Twitter DMs at scale, and an email inbox sorter — each built with Claude, then handed to Jev to run for fractions of a cent.

💡 Key Takeaways

1. Frontier + System 1 = the unlock. Claude Opus 5.5 for reasoning, Jev for decisions — computer use faster than a human, nearly free.
2. Computer use is all tool calls. "Do this step, then this step" doesn't need essay reasoning — which is why a decision model fits.
3. Jev picks, doesn't write. Under 500ms, fraction-of-a-cent decisions, no hallucinated fields (schema-constrained).
4. Build with browser-use, run with Playwright. Dynamic exploration while learning; static steps once automated.
5. The method is two-phase. Claude learns the workflow and writes a SKILL.md; Jev runs it as a choice option; Claude stays as a fallback.
6. External brains plug in. Stockfish feeds Jev the chess probabilities — the decision model doesn't have to be the smart part.
7. Iterate with Claude, not around it. When Jev fumbles a step (the marketplace message), fix it in chat, re-hand to Jev.
8. It scales to a whole agency. LinkedIn/Instagram/Twitter outreach and email sorting — all built this way.

🔗 Resources & Links

• 🐙 github.com/kevinbadi/jev-voice — Kev's Jev project (local whisper.cpp + one Jev call per command + macOS automation)
• 🐙 github.com/browser-use/jev-ultrafast — the browser-use + Jev tool stack (~20K stars)
• 🐙 github.com/microsoft/playwright-mcp — Microsoft's Playwright MCP server (~38K stars)

Source video: youtube.com/watch?v=uCZ6sxD0yhk. Companion piece to the earlier Jev-series deep dives — this one covers computer use.

⏱️ Timestamp Index