The madewithjev gallery exists to answer one question — what are people actually building with Jev, TypeSafe AI's System One decision model? — and its first month answers with numbers. It counts 838 public builds across 28 days, each linked to its source and tagged with the cost and speed its author reported. The gallery's X wave, 319 posts captured on 2026-10-15, shows the same shape at a human scale: mostly small, single-session builds that put one fast decision in front of a browser, a feed, or an agent. Everything below stays inside that gallery, and every figure is a claim its author made.
What the gallery counts
The madewithjev directory indexes 838 builds across 21 use cases, with category filters and free tools; the counts are the directory's own tallies. Its Jev Build Report puts the median published cost at $0.000065 per decision.
| Build, as listed | Reported figure |
|---|---|
| Zillow search over listings | under 20 s, $0.18 |
| Cortex computer use | 23 s down to 0.8 s |
| Jev on Cloud Run (RTX PRO 6000) | about 47 s cold start, about 117 ms end to end |
| Claude Code rule check | 348 ms |
| Wallet screen before a TGE | 40,000 wallets in 4 s |
Every figure is the author's or the gallery's; the linked source makes it checkable.
The single-session shape
Read as a sample, not a census: most builds are small enough to be one sitting. Awais's hover-explanation demo sits at about $0.006; _kayato shipped a Chrome extension that blocks spoilers; choroshin built a phishing-message detector; ericjingyang's Not Slop extension detects LinkedIn and X slop as you scroll. Charlie Hills collected twenty such builds in a single thread, each with its skills.sh link.
Starting is just as light. crespodario put Jev in a chat you can try online, with the code published; AHSeslami's first experience was a token from OpenRouter plugged into a Grok bot; BAI_AGI announced Jev live on the B.AI API; coccoinomane points at a local, open-source version. @0xRicker's Jev Engineering post gives the division of labour a name: agents create the paths, and Jev decides which survive.
What gets built first
Browsers and computer use recur more than any other category. The gallery lists a Browser Use flight search at about 7 seconds and roughly $0.004, and a browser extension that hides X posts in plain language. Cartwright's Cortex is a local MCP server where Claude plans and a fast layer does the clicking. A browser_use post has Luna planning and Jev acting in a poker game that wins. One entry runs the gallery's popular browser-control scenario on a cloud machine for AI employees. Browser agents and computer use covers why this was the first breakout category.
The decision layer for agents
The other cluster sits behind someone else's loop, returning a typed decision instead of prose — state in, auditable decisions out, as @0xChaseTM's post puts it. erkamyaman_ng checks every Claude Code reply and edit against the rules you wrote. @0xKiyoro used Jev to catch Opus 5.5's default effort dropping from high to medium, where agents without settings lost reasoning. @cyrilXBT splits the work — let Claude reason, let Jev decide — and reports the approach as 200× faster and 400× cheaper than a plain loop, plus a coding-agent memory fix called Beacon. @cesaremuszka added Jev to CodexRouter and plans a custom harness next; burakkarakann routes each Iceberg query to the right engine; and emoubarak_'s jev-check gives the pattern its crispest line: the database, not the agent, decides whether the work is done. Putting Jev in the coding loop covers the design side.
Games, markets, and comparisons
Games became the shared test bench. bubosees put three new decision models into one game of Snake — Amazon's Strands Decider 2B, Clef-flash and pplx-decider. Jev and Laya played chess in at least two posts, one arguing benchmarks are not enough and real interaction is the point. cola_runner ran Microsoft Decision-1 against Jev 1.13 in Contra with identical settings and 600 decisions each: Jev got farther, and neither cleared Stage 1. After Snake and Tank with Jev, dennis_huangbei notes, someone built a Monument Valley-style 3D game. The wave prices other vendors' APIs too — a Pokémon FireRed run on Perplexity's Decisions API cleared the Elite Four and Champion in one shot for $0.028 — and CLM-8B's release claims up to nine times the speed of Jev.
Markets attract the same shape with larger numbers attached. 1kleos1 reports a student in Japan whose Jev Polymarket bot earned $17,831 in its first night; Cygnus_DEX reports a Jev-powered AutoScalper passing testnet with an 82% average win rate on the one-minute chart; burstingbagel screened 40,000 wallets in four seconds before a TGE. Not all traffic stays: adithatipalli moved a video pipeline's decision layer from Jev to Laya.
Numbers to check, not trust
The gallery is explicit that its figures are self-published, and the sample shows the ground that covers: per-token pricing ("$0.042 per million input tokens, output free," 0xbelorix notes), a bill after free credits ran out ($5 for 123M tokens, per dsmiley411), and a customer count for the question API (16 paying customers, from AlexanderTw33ts).
Skeptical takes belong in the same record. AradhyeAgarwal argues that Jev's modeling choices lead to real failure modes; david1989_zhu warns that coverage of Vercel paid teams is not product value, since Jev judges rather than writes; code_saksham says an idea hit a wall and asks TypeSafe to fix a limit. Claims like CLM-8B's "up to nine times faster" are benchmark claims that only the source can settle.
Which is the gallery's contribution: 838 entries that each resolve to a build and a checkable number, plus a first-month pattern — decisions in front of browsers, feeds, and agents, priced per call. Whether the price holds at your volume is the number that matters; the cost of a decision is where the arithmetic lives, and the directory is where the sources do.