A week of Jev repos — routers, MCP, graphs, games

Forty-eight hours after launch, the useful question on X was no longer “what is Jev?” It was “what do you actually point it at?”

A roundup from @0xLogicrw, framed as an Awesome list for people who had just received API access, grouped early repos by the same underlying move: ask a bounded question, get a probability, let code act. A later follow-up pointed at a generated directory that claimed 130+ projects and a twelve-hour GitHub Actions sweep. Independently, hellogumbo/awesome-jev grew into a large public catalog. Counts on those lists move daily; treat them as snapshots.

The original fourteen entries are a decent map of week-one imagination:

1. jev-ultrafast — Browser Use’s indexed action-space agent.
2. typesafe-mcp — a single-binary MCP so Claude Code, Claude Desktop, and Codex can call Choice / Score / Noul.
3. jev-mcp — verify, screen, and rank tools for agents that should not spend a frontier call on mechanical checks.
4. SemDecide — Unix-shaped semantic grep / jq for pipelines and CI.
5. Jev Codex Router — per-turn model and reasoning-depth routing; the author reports about 60% cost reduction on a 237-turn replay.
6. Winnow — keep-or-drop for oversized tool results in Claude Code.
7. jev-review — first-pass review on correctness, security, reliability, then a heavier model on the flagged issues.
8. Blink — semantic codebase walk: at each directory, score which children are worth expanding.
9. neo4jev — the same idea on a graph: classify the next edge, beam-search a path.
10. jev-desktop — accessibility-tree desktop loop inside agent-desktop.
11. TypeSafe AI Playground — a Rust CLI of small experiments (PHI detection, comment review, tone).
12. Prism — the thread’s closest trading example: Jev scores toxic flow, pressure, and mean reversion, defaulting to shadow/advisory rather than placing orders.
13. OneVOneJev — an FPS-style demo with move / aim / fire decisions around 9 Hz.
14. safer-with-jev — the URL in the original roundup: Jev as a router in front of larger chat models. A related Neon write-up lives at typesafe-on-neon.

The roundup’s closing line is the design rule. Use Jev for “should we, which one, what score, what next.” When you need a patch, an essay, or a long plan, call a language model.

Savings and latency numbers inside those READMEs are author-reported. The more durable signal is how quickly people converged on the same four verbs.