The Model That Does Not Speak: 7 Free Repos That Speed Up Claude Code

Amid the chatter about AI chatbots I heard a strange sentence: a model that never speaks makes decisions 40 to 200 times faster than the ones that do. TypeSafe calls it Jev, a System One model that produces probabilities instead of sentences, with responses in the 70-500 millisecond band, input priced at $0.042 per million tokens and output free. To make that abstract promise concrete, the host opens seven free repositories that plug into Claude Code, and states the thesis up front: let the smart model plan, let the fast model decide.

The box that turns into UI as you type

The first stop is a demo where the box you type into reshapes itself around your intent. One line becomes an event card or a checklist; the next turns into a timer or a color picker; bill splitting, polls and converters follow the same pattern, because the box estimates your intent probabilistically and builds the interface instantly. The examples flow: a color search resolves in a blink, dinner plans for Friday land on the calendar, a 25-minute countdown and a simple division appear without waiting. On its own it looks like a toy, but the host's example scales it up: imagine an order box that buys automatically when a price drops 5 percent, and suddenly milliseconds turn into money. The instant UI generation he describes is reportedly what major developers chase in calendar and messaging products.

The second repository puts the browser under voice control, and the difference hits in the first second. Classic voice assistants leave several seconds of dead air between command and action; here the model anticipates intent while the words are still leaving your mouth. In the demo a navigation command opens the page at once, a click-first-link instruction executes without delay, and a spoken place name drops results onto the screen. The host admits his own recording clashed internal audio with video capture, so he shows someone else's demo, but adds that he tested the tool himself and found it lightning quick . Given how laggy voice commands have always felt in Hermes-style agents, the gap is easy to appreciate.

The third repository promises to rewrite computer use, and it brings numbers. Conventional agents take a screenshot and predict where to click, which costs around 5 seconds per command. The Jev-based tool predicts the click target from the command itself and cuts the time to 0.3 seconds. The shared table is bold: 119x cheaper per input token, 155x cheaper per decision, 0.13-0.38 seconds against 5.2 seconds, a 14-40x speedup. Developer Aaron's page sits at only 955 stars, which the host reads as youth rather than weakness. The demo that breaks down a TechCrunch page and extracts pricing shows why reading the backend beats re-shooting the screen.

Using the computer without screenshots

The fourth example is pedagogical rather than practical, by the host's own admission: a blocker that reads pages live and deletes ad slots on sight. Every site entered gets scanned and cleaned, yet nobody would install this for speed while mature blockers exist. Still, its place on the list has a logic: an AI reading a page and answering a binary question in milliseconds is everyday proof of the architectural shift. The host says so honestly and moves on to the heavy hitters.

The fifth repository is, in my judgment, the most useful stop in the video: a set of MCP judgment tools that dock into Claude Code. Verify, screen, find and decide check claims against evidence, filter content before it enters context, pick the best candidate by meaning, and settle bounded alternatives. Every result returns with a typed probability at a fraction of a cent: in the live demo a pricing verification costs $0.000023 and lands in 150-500 milliseconds. Cheap mechanical checks that frontier models skip or overcharge for move to Jev. The key point is that the harness never changes: you stay inside Claude and call Jev only at the moments it is good at.

The sixth chapter wires everything into automation lines, with examples from the host's own business. Agency candidate screening becomes three probability questions: the right person, a company of interest, criteria fit. The output feeds back as a ranked probability list and outreach runs automatically. In the trading-bot example the decision lands in 400-500 milliseconds; scanning news flow, it issues a buy verdict at 91 percent under preset rules. The thesis is crisp: use a probability filter over thousands of records instead of reasoning over each one, and reserve expensive inference for the narrowed list. Where speed is directly money, as in high-frequency flows, this architecture shines brightest.

Live scoring and the idea catalog

The seventh stop is playful but eye-opening: a live nonsense meter scoring a broadcast debate sentence by sentence. Each line gets an instant credibility score; the host's first thought is prediction markets, where faster judgment means faster positions. The generalization is more interesting: measuring live which webinar sentence moves sales, which phrasing closes clients, indexing consumer reactions in live shopping streams. Abstract today, these ideas could become tomorrow's reporting tools once live video analysis gets cheap. Impractical for now, but a door worth peeking through.

For the closing the host opens what he counts as the seventh resource: a catalog of what people build with Jev, browsable by build type. Agents and browsers, games and real time, trading and markets, content and growth, each card carrying cost and time notes. His suggested method is clever: hand the site to Claude Code together with your own context file and ask which idea fits your current business or could grow into a new venture. The last line reads like the video's summary: Jev belongs beside a language model, not instead of it; it accelerates and cheapens . The Opus 5.5 praise stays a personal view, and since no official release info is independently confirmed, it deserves a cautious read.