Vibe-Trading
A personal trading agent that connects market data, research workflows, and backtesting.
Per-dimension scores and reasoning
Evidence shows: SECURITY.md explicitly states that generated backtest code runs locally and the subprocess environment is restricted, not forwarding LLM provider keys, API server tokens, shell-tool opt-ins, broker trading secrets, etc., reflecting least privilege. User confirmation: README mentions live risk-increasing actions are mandate-gated and audited, but lacks specifics. Data flow transparency: README mentions hash manifests and hash-chained audit ledger, but not detailed data flow. Sensitive data handling: SECURITY.md mentions not forwarding sensitive keys, but not storage/encryption details. Dependency security: pyproject.toml shows version bounds, CI has hash-lock check, but no vulnerability scanning evidence. External effects: README mentions live trading actions require authorization, but not detailed. Rollback: README mentions session/run state, but no explicit rollback mechanism. Source attribution: README and SECURITY.md clearly identify official channels and impersonation warnings, but no full contributor list.
Evidence shows: README news section documents multiple correctness fixes (price validation, interval handling, data source fixes), indicating consistency focus. Dependency availability: pyproject.toml lists many dependencies with optional extras, CI verifies installation, but no guarantee for all. Failure messages: README mentions error handling like 'fail-closed tool arguments', but no detailed error message examples.
Evidence shows: README describes multiple use cases (backtesting, research, live trading) and multi-language docs, indicating broad audience. Capability boundaries: README mentions read-only tools and gating for live actions, but not exhaustive. Trigger precision: README mentions price validation and interval handling, but not detailed trigger conditions. Environment fit: README mentions OS and Python version support, but not all environment requirements.
Evidence shows: README has clear structure (News, Features, Quick Start, Examples, API/MCP, Roadmap), good information architecture. Install notes: README provides pip install command, but not detailed steps. Naming stability: project name and package name consistent, but no naming conventions. Examples and FAQ: README provides examples, but no FAQ. Known limitations: README mentions some (Python version upper bound), but not comprehensive. License: LICENSE file is MIT, standard. Versioning/changelog: README news section records changes, but no formal changelog. Maintenance responsibility: README mentions contributing guide, but no explicit maintainers.
Evidence shows: README describes various outputs (backtest reports, risk analysis), but no specific output formats. Marginal value: README mentions many features (multi-market support, quant library), but no comparison with alternatives. Cost-benefit: README mentions free data sources, but no performance benchmarks.
Evidence shows: README news section references many PRs and issues, providing traceability. Cross-source corroboration: README mentions multiple data sources, but no independent verification. Fact-inference separation: README mentions 'marks what a source does not state instead of inferring it', indicating focus on distinction.
- Static review cannot verify actual runtime behavior; all scores are inferred from file contents.
- Live trading features involve high risk; use with caution and ensure understanding of authorization and audit mechanisms.
- Many dependencies; recommend regular vulnerability checks and updates.
- No detailed rollback mechanism provided; recommend backup before critical operations.
What does this agent do, and when should you use it?
Vibe-Trading is a personal trading agent for quantitative research and trading workflows, with a CLI, FastAPI backend, React Web UI, and MCP server. Its market-data loaders, research tools, factor capabilities, and portfolio backtests move an idea from hypothesis through signal generation and evaluation. Outputs include run records, strict-JSON validation, backtest attribution, and risk_xray.json and risk_xray.md for portfolio backtests. It can use local CSV, Parquet, or DuckDB data alongside multiple market-data and broker integrations; live-trading paths have authorization, mandate, and risk boundaries. Docker deployments can preserve user data, sessions, skills, shadow accounts, and broker configuration in named volumes.
The local loader reads OHLCV data from CSV, Parquet, or DuckDB files, while the loader registry fetches and normalizes market data. run_research_autopilot and generate_backtest_config connect hypothesis → research goal → backtest; scaffold_signal_engine creates a contract-compliant signal engine, and link_autopilot_backtest feeds metrics back into the hypothesis workflow. alpha bench --strict applies same-universe random controls and an out-of-sample gate, while portfolio backtests emit risk_xray.json and risk_xray.md. The MCP server exposes market-data, factor-analysis, and research-related tools; the CLI includes vibe-trading provider doctor, vibe-trading resume, and connector commands. The system also uses fallback chains for missing symbols and validates prices, returns, OHLC data, intervals, and strict JSON in multiple paths.
- A quantitative researcher who needs to turn an investment hypothesis into a signal engine, backtest, and traceable metrics can use the Research Autopilot flow.
- An individual investor with CSV, Parquet, or DuckDB price data who wants to backtest without relying solely on hosted market data can use the local loader.
- A research team comparing factor or portfolio strategies and checking out-of-sample constraints and risk exposure can run alpha bench --strict and inspect risk-x-ray artifacts.
- A developer who wants to connect the same trading-research runtime through chat, a terminal, or a protocol interface can use the React Web UI, CLI, or MCP server.
- A strategy developer researching US, Hong Kong, A-share, Indian, Korean, or USD-M perpetual markets can use the documented market engines and data routing.
- A user connecting a brokerage account while needing explicit live-trading controls can use connectors, noting that some are read-only and live actions are constrained by mandate and safety controls.
What are this agent's strengths and limitations?
- Research Autopilot explicitly connects hypotheses, signal engines, backtests, and metric feedback instead of only generating trading commentary.
- The backtesting and data layers include targeted correctness controls: strict JSON, OHLC integrity checks, out-of-sample gates, causal rebalancing, and risk-x-ray artifacts.
- One runtime is exposed through CLI, Web UI, FastAPI, and MCP, with a local data bridge and market-data fallback chains.
- Live-trading features have authorization, mandate confirmation, exposure limits, kill-switch, and audit boundaries; some connectors, including Trading 212, remain explicitly read-only.
- Model-dependent research requires LLM provider credentials, but the supplied material does not include complete provider-configuration instructions.
- Data sources, broker connectors, and optional premium data vary by account, key, region, and dependency; not every connector can place orders.
- The project has recently shipped recurring fixes across loaders, return calculations, provider adapters, and backtest correctness, so adoption should include validation against the intended market and strategy.
- Its Docker, Web UI, messaging, MCP, brokerage, and multi-market surface creates more operational and permission-management work than a single-purpose backtesting library.
How do you install or deploy this agent?
Python 3.11+ is required. Install with: pip install -U vibe-trading-ai. Then run vibe-trading provider doctor to inspect a redacted provider, model, package, and proxy diagnostic snapshot. Running model-dependent research also requires credentials for a selected LLM provider; the supplied material does not provide exact copyable configuration fields or a full startup command.
How do you use this agent?
Start with vibe-trading provider doctor to validate the provider environment. The documented command surface includes non-interactive vibe-trading run, vibe-trading resume <session-id>, alpha bench --strict, and connector; after provider credentials are configured, these can run research, resume sessions, evaluate alphas, and manage connectors. For local data, place OHLCV data in CSV, Parquet, or DuckDB and use the local loader. The supplied evidence does not include a complete argument example for vibe-trading run, so a first research command should not be invented.