vibe-trading

What it is

A multi-agent finance research workspace, self-hosted on the kulify VPS. Open-source from HKUDS — I run it for myself rather than building it. Natural-language → trading strategy / research / backtest pipeline, with 74 finance skills, 29 multi-agent swarm presets, 7 backtest engines, and 6 free data sources with auto-fallback.

Why I host it

Vibe-Trading is one of the most interesting reference implementations of multi-agent finance work I’ve found — explicit skill registries, a clean SSE-based event bus between agents, a backtest engine that generalizes well beyond finance. Running it gives me a daily workspace plus a live system to study patterns in.

The deployment itself was instructive: Vibe-Trading uses shell tools as a first-class capability, which means containerization is the sandbox. Single-user bearer auth, DeepSeek as the model provider, three persistent volumes for runs / sessions / .vibe-trading config. The first deploy hit Coolify’s healthcheck-vs-python:3.11-slim mismatch (slim images don’t ship curl/wget), which became one of the captured Coolify gotchas in the kulify wiki.

What I’ve taken from it

• The persona pattern — agents as named experts with their own deterministic analysis pipelines plus LLM reasoning, all emitting the same structured output schema. Generalizes far beyond trading.
• The skills system — progressive disclosure for tool-using agents, where the registry exposes a tiny core and the agent loads more on demand. Pattern documented in the kulify wiki, queued as a lift into other projects.
• The agent self-modification boundaries — Vibe-Trading lets agents write their own skills, edit files, and run bash. The trust zones (user skills, run-dir workspace, system-via-bash) are the first concrete reference I’ve seen for “what does sandboxing actually look like for an agent that codes.”

Notes

This is a third-party project. I deploy it, study it, and lift patterns from it — not a project I built or maintain. Credit: HKUDS.