Most AI-driven trading demos either answer from model priors or expose blunt live-execution hooks. Vibe-Trading takes the opposite stance: treat the agent as a research-first orchestrator that produces testable artifacts (code, backtests, run cards, reports) and only bridges to execution with explicit, mandate-gated connectors and filesystem-level kill switches. That design pushes automated strategy discovery toward reproducibility and auditability instead of opaque autonomous orders.
What Sets It Apart
- Tool-first agent surface: exposes 36 MCP tools and a full MCP server/client flow so external agents (Claude Desktop, OpenClaw, etc.) can call its research backends or embed its skills. This makes integrations deterministic and scriptable rather than ad-hoc LLM prompts.
- Rich finance library + prebuilt alphas: bundles 77 finance skills and an Alpha Zoo (452 alphas across Qlib/alpha101/GTJA/academic), plus cross-market backtest engines (A/HK/US equities, crypto, futures, forex) and multi-source data loaders (yfinance, AKShare, mootdx, CCXT, etc.).
- Research-first safety model: research artifacts (run_card.json, run traces, persistent memory) come first; broker connectors are opt-in and guarded by committed mandates, pre-trade gates, and audit ledgers — trading features are explicitly experimental and gated behind user consent.
- Shadow Account & reproducibility: built-in trade-journal parsing, shadow-account comparisons, and run cards to reproduce and audit how a proposed rule set would have behaved versus real trades.
Who It's For (and trade-offs)
Great fit if you build, test, or audit systematic strategies and want an agent that generates reproducible research artifacts: quant researchers, quant-savvy PMs, developer-first trading desks, and AI integrators who need an MCP-capable skillset. It’s also useful when you want local/offline LLM workflows (Ollama) or multi-agent orchestration.
Look elsewhere if you need a turnkey retail trading product or a non-technical managed service: the system assumes familiarity with CLI/Web devops, LLM provider configuration, and backtesting concepts. Live trading features are experimental — use paper/demo connectors and read the mandate/kill-switch docs before authorizing any live paths.
