An agent builder disguised as a mobile jumping game

What it does

Jumping-Agent is a browser-based construction kit that lets users assemble AI agents by hopping between 3D platforms instead of wiring nodes on a flat canvas. A Three.js frontend translates workflow steps into spatial jumps; once the structure is built, a Python orchestrator generates the actual agent workspace and can bridge it to WeChat for conversational access. The whole stack is aimed at novices who want to see an agent run without first learning graph theory.

The interesting bit

The project treats workflow topology as level design: you press, hold, and jump to stitch together sequential, routing, or parallel steps, which is either a brilliant UX hack or a very elaborate way to avoid drawing arrows. Underneath the game layer, a ReAct-based back_agent fills in and rewrites the generated skeleton code so the visual structure actually becomes a runnable Python project.

Key highlights

• Three.js “jumping” interface replaces traditional node-and-wire editors for tablet users.
• WeChat integration via QR-code binding turns the built agent into a chatbot you can message directly.
• ReAct agent back-end auto-completes and rewrites the scaffolded code based on frontend inputs.
• Supports sequential, routing, parallel, and a handful of other workflow templates out of the box.
• Solo-developed; the author openly notes it is still rough around the edges.

Caveats

• Only seven workflow templates exist so far, and the README admits the jump-platform orchestration can be unstable.
• Setup currently requires manually launching three separate services across two runtimes; the promised CLI is still on the roadmap.
• WeChat integration relies on a specific iLink connector, so the out-of-the-box chat experience is tied to that ecosystem.

Verdict

Worth a look if you are experimenting with no-code agent builders or need a WeChat-native deployment path, but hardcore workflow engineers will probably miss the precision of a traditional node editor.