// the find
brokermr810/QuantDinger
AI quantitative trading platform for crypto, stocks, and forex with backtesting, live trading, market data, and multi-agent research.vibe-trading ,trading-agents,ai-trader,ai-trading
QuantDinger is a self-hosted quant trading stack that connects AI-assisted research, Python strategy authoring, server-side backtesting, and live execution across 10+ crypto exchanges plus IBKR, MT5, and Alpaca — all in one Docker Compose deployment. It targets independent quants and small teams who want to own their infrastructure and exchange credentials rather than hand them to a black-box SaaS. The MCP server on PyPI makes it usable directly from Cursor or Claude Code.
The dual strategy runtime is the best design decision here: IndicatorStrategy (vectorized dataframe signals, good for research and charting) and ScriptStrategy (event-driven on_bar with explicit orders, good for live alignment) are genuinely different mental models and keeping both in the same codebase avoids the usual research-to-production translation pain. The agent safety model is thoughtful — paper-only by default with a double-unlock requirement (token flag AND server env var) for live trading, plus append-only audit logs on every agent call; this is not just a disclaimer, it is enforced in the execution path. The test suite has real depth: exchange-specific tests, fixture-based contract tests for order and position shapes across venues, fill recovery tests, grid reconciliation tests — the kind of coverage that only gets written after getting burned in live trading. Multi-arch GHCR images with a one-line curl install and no Node.js requirement for the frontend is genuinely frictionless.
The frontend (QuantDinger-Vue) is not Apache 2.0 — it uses a 'source-available' license, which the FAQ mentions once and the main README mostly glosses over; operators building commercial products need to read that license carefully before assuming this is fully open-source. Strategy code runs through safe_exec.py on the server, but the README says nothing about the sandboxing mechanism — in a multi-tenant deployment this is a critical attack surface and the silence is a red flag. Schema management is a single init.sql file run idempotently at boot, which works fine at initial setup but leaves the upgrade story completely unaddressed; when you need to ALTER a column after the first deploy there is no Alembic or equivalent, so schema drift is your problem to solve. The AI strategy generation feature is presented uncritically — LLMs produce plausible-looking Python that backtests acceptably and falls apart in live markets; the README should say this plainly rather than treating generated code as a starting point without caveats.