Vibe-Trading

Vibe-Trading is an open-source trading agent built for developers and quant researchers who want to describe a trading idea in plain language and have an agent turn it into a backtested, executable strategy. It combines a FastAPI backend, a React frontend, and an MCP server so the same trading toolkit can be driven from a CLI, a web UI, or any MCP-compatible AI client. It's aimed at people comfortable with Python who want programmatic control over data sourcing, factor research, backtesting, and order execution, not at people looking for a plug-and-play robo-advisor.

The project ships as a pip package (vibe-trading-ai) and is under active development, with frequent PRs covering backtest correctness, security hardening, and new market/data integrations. It supports multiple LLM providers (OpenAI, DeepSeek, Zhipu GLM, NVIDIA NIM, and others via OpenAI-compatible gateways) and multiple market data sources, including free providers and an optional paid tier (QVeris) for broader coverage.

Key features

• Multi-agent architecture: the agent decomposes a trading request into research, factor development, backtesting, and execution steps rather than running as a single monolithic prompt loop.
• Alpha Zoo factor library: hundreds of pre-registered quantitative factors (academic and broker-derived) that strategies can reference directly.
• Strategy Development Manager: a skill that converts papers or research notes into registered factors/strategies, tracked through an active → monitoring → decayed → disabled lifecycle with IC/Sharpe decay monitoring.
• Backtesting engine: causal, order-independent rebalancing, terminal close cost accounting, fill-derived turnover reporting, and exposure caps, aimed at avoiding common backtest correctness bugs like look-ahead bias.
• Multi-market data support: equities, crypto (including Binance USD-M perpetuals), and region-specific engines such as an Indian equity engine (NSE/BSE) with local cost and settlement rules.
• MCP server: exposes trading tools over Streamable HTTP, so any MCP client can call backtesting, data, and order-placement functions.
• Shadow Account mode: lets you run and evaluate strategies without committing real capital, with a separate confirmation step required before a live mandate is set.
• IM channel integrations: multiple messaging adapters for receiving alerts or interacting with the agent outside the web UI.
• Security-hardened deployment: Docker images are digest-pinned and built multi-stage, the backtest sandbox blocks network/subprocess/eval calls, and API endpoints carry auth and rate limiting.

Ideal use cases

Vibe-Trading fits developers building or automating a personal or small-team quant workflow: researching factors, backtesting rule-based or LLM-assisted strategies, and wiring an agent into a broker or data feed via MCP. It also suits people who want an AI agent to draft and test a strategy from a natural-language description before manual review.

It's not a fit if you need a regulated, institution-grade execution platform with guaranteed uptime and compliance certifications, or if you want a no-code product with a hosted backend you don't have to run yourself. It also isn't designed for high-frequency trading; the backtesting and agent-orchestration layers assume research and swing/position-style workflows rather than microsecond execution.

Installation

The package is distributed on PyPI and requires Python 3.11 or later.

pip install vibe-trading-ai

To update to the latest release:

pip install -U vibe-trading-ai

After installation, configuration (LLM provider keys, data provider keys, broker credentials) is stored centrally in ~/.vibe-trading/.env, and can be set through either the CLI onboarding flow or the Web Settings UI. The backend runs on FastAPI and the frontend on React 19; a Docker Compose setup is also available for containerized deployment with a hardened, read-only root filesystem and dropped capabilities. Check the project docs at vibetrading.wiki for provider-specific setup (LLM keys, market data credentials, and broker connections) since these vary by which markets and data sources you plan to use.