10 个可直接上手的 Jev 项目：桌面、交易、马里奥

ggwp @ggwplabs原推发布于 2026.10.02

在 X 上打开原推

Jev already has 10 projects you can build on. You don’t need another prompt pack. Desktop agents. Trading experiments. A model playing Mario directly from emulator state. The interesting part is what these projects have in common: Jev chooses the next action. The surrounding code decides what it can actually execute. Here are 10 starting points: 1. agent-desktop Give an agent access to real buttons, menus and input fields through the operating system’s accessibility tree. Observe the interface, select an action, execute it, check the result. 2. typesafe-mario Let Jev play Super Mario from structured emulator state. Instead of interpreting screenshots, it receives game data and chooses its next move. 3. jev-drone A drone-control experiment in MuJoCo. Jev participates in the decision loop. The flight controller still handles low-level stability. A simulation, not proof you should hand it a real drone. 4. OneVOneJev A browser-based FPS arena. Turn movement, aiming and firing into decisions the model can select from the current game state. 5. jev-trader An experimental trading loop on Monad using Kuru’s MON-USDC market. Read the order book, request a decision, pass it through execution constraints. Start with observation or paper testing. A fast decision is not a profitable strategy. 6. Prism A liquidity-management agent with a different role for Jev. Evaluate market conditions and feed those judgments into the strategy instead of letting model output become an unrestricted order. 7. neo4jev Let Jev navigate a Neo4j graph. At each node, choose between neighbouring relationships and continue the search. Useful when the problem is deciding where to look next. 8. jev-curate Filter training data from JSONL and Parquet. Use typed judgments to evaluate quality, relevance and risk before accepting a row. 9. Canny Make coding agents back up “done” with evidence. Deterministic hooks enforce the checks. Jev provides advisory judgments. The agent’s confidence is not the same thing as passing tests. 10. killmyidea Feed it a startup idea and get a structured verdict: KILL, FIX or SHIP. Use the verdict to challenge your assumptions, not replace customer research. The build pattern worth copying: → expose the current state → define the allowed choices → ask Jev for a decision → validate it in code → execute within limits → observe what actually happened Start with one bounded task. Log the decisions. Test the failure cases. Keep a fallback for uncertainty. These repos give you somewhere concrete to start instead of another blank agent loop. Save this for your next Jev build.

这条是引用着另一条写的：x.com/ggwplabs/status/2096584388288852295