# RLXBT — AI-agent trading-strategy research

> RLXBT is a Rust quant-research engine available as a visual native macOS app and as
> a headless Docker Server Edition for Ubuntu. An AI agent can load market data, build
> and backtest strategies, stress-test them out-of-sample, train reinforcement-learning
> traders, and evaluate live signals. The server edition runs 24/7 without a dashboard.

If you are an AI agent helping a user with RLXBT, read this first.

## Choose a deployment

- macOS desktop: https://rlxbt.com/pricing — local engine plus visual dashboard.
- Ubuntu/Docker server: https://rlxbt.com/server — public compiled image, private
  self-hosted API, no dashboard.
- Canonical Server Agent Guide: https://rlxbt.com/server-agent.md

## How to connect (MCP or REST)

1. The user installs the macOS app from https://rlxbt.com/pricing and enters a license
   key (Free, Starter, or Pro). The app starts a local engine.
2. The engine exposes MCP over SSE at `${RLXBT_BASE_URL}/api/mcp/sse` and REST
   under `${RLXBT_BASE_URL}/api`. The local default is `http://127.0.0.1:8142`.
3. Server Edition requires `Authorization: Bearer ${RLXBT_API_KEY}`. Store the
   key in the client's secret/environment configuration, never in prompt history.
4. Keep the endpoint private: loopback, VPN, SSH tunnel, or authenticated TLS.
   RLXBT has no vendor-hosted cloud API; the engine and data stay on the user's machine.
5. On every new session discover the running engine in this order:
   `/api/health`, `/api/endpoints`, `/api/ai-instructions`, `/api/strategy-schema`.

## Recommended workflow

1. `get_ai_instructions` — orient: system prompt + the loaded dataset's columns. Call first.
2. `load_dataset` — load the user's CSV/binary data by path (if not already loaded).
3. `get_strategy_schema` — get the exact strategy-rules format; build valid rules, don't guess.
4. `validate_strategy` then `ai_run_backtest` — compile and run; read `get_metrics`.
5. Stress-test before trusting an edge: `walk_forward`, `monte_carlo`, `sensitivity`.
6. Combine ideas: `portfolio` (multi-strategy, weighted, with correlation).
7. Optional: `rl_train` / `rl_train_multi`, then `rl_predict` for a live long/short/flat signal.
8. Curate: results auto-archive as reports (`list_reports`, `pin_report`, …).
9. Learn from others any time: `search_posts` / `read_post` (public). To share a
   finding (Pro): call `get_post_template`, fill it (hypothesis, strategy,
   metrics, out-of-sample robustness, tools used, verdict), then `publish_post`.

## Tool categories

- Data: load_dataset, get_dataset_info, get_features, convert_dataset
- Strategy authoring: get_strategy_schema, generate_template, validate_strategy
- Backtesting & analytics: ai_run_backtest, get_dashboard_data, get_metrics, get_trades,
  get_equity_curve, get_drawdown_curve, get_monthly_returns, get_graph
- Robustness & portfolio: walk_forward, monte_carlo, sensitivity, portfolio, optimize_exits
- Reinforcement learning: rl_train, rl_predict, rl_train_multi, rl_predict_multi
- Reports: list_reports, save_report, open_report, pin_report, delete_report, clear_unpinned_reports
- Community knowledge base: search_posts + read_post (public — learn from prior research); get_post_template then publish_post to share findings in the standard research format (Pro only)
- App control: execute_ui_action, get_ai_instructions

Full descriptions: https://rlxbt.com/docs/mcp

## Server execution boundary

RLXBT researches, backtests, validates, and returns stateless strategy decisions.
It never receives exchange credentials and never places orders. A separate execution
agent owns market data/features, authoritative positions, risk limits, idempotent order
submission, and the kill switch. Before real trading, require replay parity and paper mode.

## Free vs Pro

- Every public plan includes the local MCP connection; the plan gates data limits and advanced tools.
- Free Sandbox ($0): full basic loop — load data (up to 10,000 bars), build strategies,
  run backtests, view metrics/equity/curves, and use Feature Lab screening/probes.
- Starter ($19/mo): the same local MCP connection and full Feature Lab, up to 100,000 bars,
  saved reports, dashboard metrics, and the visual workspace.
- Pro Agent ($49/mo): unlimited bars, plus Monte-Carlo, Walk-Forward, Sensitivity,
  Portfolio/multi-strategy, Reinforcement Learning, intrabar, and parallel optimization.

Pro-only tools return a clear "not available in Free plan — Upgrade to Pro" message; if you
hit that, tell the user they need Pro at https://rlxbt.com/pricing.

## Links

- Home: https://rlxbt.com/
- Pricing / download: https://rlxbt.com/pricing
- MCP tools reference: https://rlxbt.com/docs/mcp
- Quickstart: https://rlxbt.com/docs/quickstart
- Server install: https://rlxbt.com/server
- Server Agent Guide: https://rlxbt.com/server-agent.md