Vibe-Trading: AI Trading Agent for Backtesting & MCP
/Vibe-Trading
Vibe-Trading is an open-source agent that hooks an LLM up to real market-data feeds, nine backtest engines, and broker connectors through MCP tools, instead of letting the model guess prices from memory. Reach for it if you're a Python developer who wants an agent that can run and check its own backtests; skip it if you need a stable system for live order execution, since the changelog shows core math bugs still getting fixed after release.
Your Personal AI Trading Agent
Vibe-Trading is a Python trading library and AI agent that GitHub tags as ai-agent, algorithmic-trading, backtesting, fintech, llm, mcp, multi-agent, and quantitative-finance. One command connects an LLM to market data, nine backtest engines, technical-indicator and options-analytics tools for financial modeling, and an MCP server, so the agent runs quantitative research instead of describing it from memory.
Core Capabilities for Quant Research
- ✓Nine backtest engines cover multiple markets, with a Korea equity engine (KRX: KOSPI/KOSDAQ) added as the ninth, using execution-time price bands and a 2026 0.20% transaction tax on KRX trades.
- ✓A technical_indicators tool runs RSI, MACD, Bollinger Bands, SMA, and EMA calculations against the same data loaders used elsewhere in the project.
- ✓Options-analytics tooling works out expiry P&L extrema, exact breakevens, and spot-versus-implied-volatility scenarios through both the agent and MCP interfaces.
- ✓Four read-only institutional data tools cover SEC 13F filings with quarter-over-quarter position diffs, ETF constituent look-through, event-contract implied probability, and arXiv/OpenAlex paper search.
- ✓Six institutional research commands — /comps, /dcf, /attrib, /memo, /earnings, /screen — plus investor-lens skills and scheduled-research templates.
- ✓A local sentiment tool scores arbitrary text and pulls the crypto Fear & Greed Index, no API key required.
- ✓Persistent agent memory has an opt-in lifecycle — quality scoring, Ebbinghaus decay, archive-only garbage collection — that stays off unless you turn it on.
- ✓Every portfolio backtest writes risk x-ray artifacts (risk_xray.json / risk_xray.md) for risk management, reporting concentration, volatility, and drawdown in one file.
Applications for Financial Professionals
- •Backtesting a rule-based strategy across the nine supported markets, including the new KRX Korea engine, before committing capital.
- •Pulling SEC 13F institutional holdings or ETF look-through data into a due-diligence memo with the /comps or /memo commands.
- •Screening a large stock universe when there's no single best answer — Vibe-Trading now returns a shortlist instead of stalling out.
- •Wiring broker or exchange data into agent-driven research through the OpenBB Workspace bridge or the MT5/Binance/OKX connectors.
- •Scheduling recurring research jobs with IANA timezone support so a cadence survives daylight saving time changes.
Getting Started with Vibe-Trading
Documented only in fragments. The v0.1.12 release notes give `pip install -U vibe-trading-ai` for a pip install/upgrade, and a Docker path exists with a hash-locked dependency install that CI now checks. Application state — sessions, runs, swarm runs, uploads — lives under `~/.vibe-trading`, relocatable via the `VIBE_TRADING_HOME` environment variable. The README's own navigation points to a separate Quick Start page for the full walkthrough (API keys, MCP wiring, first agent run), which isn't included in what's available here — so treat the step-by-step setup as not clearly documented in this source.
Strengths
- ✓Nine backtest engines span multiple markets, including Korea's KRX added most recently, rather than locking you into one asset class.
- ✓Bugs get fixed in public with specific before/after numbers in the changelog — a 4.2× fundamentals understatement and a 47-percentage-point A-share return error are both named and dated.
- ✓The new institutional data tools (13F, ETF look-through, event contracts, papers) are read-only, which limits what a misbehaving agent could actually do.
- ✓CLI, MCP server, and a five-locale web UI ship as one project rather than three separately maintained pieces.
Current Status and Considerations
- △The README carries an active security warning: an X account (VibeTrading_HKU), a Virtuals project (101845), and a specific token contract are impersonating the project — Vibe-Trading says it has never launched or endorsed a token.
- △The changelog's pace of correctness fixes — a 4.2× fundamentals understatement, a 47-percentage-point A-share pricing error, a fabricated day-zero options P&L of up to +93% of premium — means numbers the agent produced before a given patch date could have been wrong.
- △A breaking MCP 2.0 upgrade is still unmerged pending a full lock/runtime migration, so the MCP surface isn't finished changing.
- △Step-by-step setup beyond the pip/Docker commands isn't documented in what's available here — you'll need the separate Quick Start page.
Comparing Trading Agent Frameworks
Frequently Asked Questions
Vibe-Trading's changelog shows frequent post-release fixes to core calculations — a 4.2× fundamentals understatement and a 47-percentage-point A-share return error were both patched shortly before this review — which argues for treating it as a research and backtesting tool rather than a live-trading system without your own verification layer.
Vibe-Trading documents nine backtest engines, with Korea equity (KRX: KOSPI/KOSDAQ) added as the ninth, plus changelog references to MT5, Binance, OKX, Tiger, Alpaca, Shoonya, Longbridge, and akshare connectors, an OpenBB Workspace bridge, and options and USD-M perpetual futures analytics.
Vibe-Trading's changelog lists direct-OpenAI reasoning_effort support, the temperature-deprecating Claude models opus-4-7, opus-5, and sonnet-5, and the zai provider on glm-5.1; its settings page also discovers configured-provider models automatically.
Vibe-Trading is released under the MIT license, per its GitHub repository metadata.
Vibe-Trading's README contains an active warning: the X account VibeTrading_HKU, alongside Virtuals project 101845, and a specific token contract are not official project assets. The team states it has never launched or endorsed a token or memecoin, and users should not connect a wallet or sign anything tied to them.
Vibe-Trading pulls from what the changelog calls 24 data sources, with per-source annualization coverage tested in CI, and applies corporate-action adjustments — Tushare A-share closes, for instance, are now adjusted around ex-dividend dates after a fix closed a return-calculation gap of up to 47 percentage points.
The problem it solves
Point an LLM at a stock ticker in ordinary chat and you get a hallucinated price or a stale training-data figure — there's no live OHLC data, no backtest to check a strategy's plausibility, and no place to run options math or SEC filings against numbers the model can trace back to a source. Vibe-Trading wires that missing tool layer directly into the agent instead of leaving it to a general-purpose chat interface.
How to use
The README frames Vibe-Trading as 'one command' to give an agent trading tools, reachable through a CLI — `vibe-trading run` covers non-interactive sessions — and through an MCP server for programmatic access. The changelog references institutional slash commands (`/comps`, `/dcf`, `/attrib`, `/memo`, `/earnings`, `/screen`), a `connector` CLI that loads broker credentials from `~/.vibe-trading/.env`, and an `alpha bench --strict` mode for out-of-sample strategy validation. A web UI adds a Scheduled page, in five locales, for recurring jobs. The full command reference sits on the linked Docs/Quick Start pages rather than in the README excerpt reviewed here.
Who should try it — and who should skip
Reach for Vibe-Trading if you're a Python developer or quant who wants an agent that calls real backtesting and market-data tools instead of guessing numbers, and you can tolerate a project still shaking out bugs in its core math. Skip it if you need an audited system for live order execution, or if tracking a changelog of specific fixes to trust a given number isn't something you want to do.
Related repositories
Still deciding about Vibe-Trading?
One click hands the question to an AI along with this page — see what it says about Vibe-Trading.