Agent Capabilities

When you connect an agent over MCP, it gets 30 tools — the entire backtester surface. This is everything your AI can do for you. No code; you ask, it calls.

Point your agent here
http://127.0.0.1:8142/api/mcp/sseSSE transport · works with Claude Desktop, Cursor, and any MCP client
🗂️

Data

4 tools
load_dataset

Load a CSV or binary .chunk dataset by path (mmap for 10GB+). The engine reads it natively.

get_dataset_info

Stats for the loaded dataset: rows, time range, price range, available indicator columns.

get_features

List the available columns / indicators (features) in the current dataset.

convert_dataset

Convert a CSV into a fast binary .chunk (bars) or .fchunk (features) for huge data.

✍️

Strategy authoring

3 tools
get_strategy_schema

The JSON schema for strategy rules — so the agent builds valid logic, not guesses.

generate_template

A starter strategy template (rules or graph) to build from.

validate_strategy

Parse & compile a ruleset with the engine's own parser before running — catches errors early.

📊

Backtesting & analytics

8 tools
ai_run_backtest

Run a strategy backtest in real time; every run is auto-archived as a report.

get_dashboard_data

Full result of the last backtest: metrics, equity curve, trade statistics.

get_metrics

Performance metrics: total return, Sharpe, Sortino, max drawdown, profit factor…

get_trades

The trade list (filter by side / limit / offset).

get_equity_curve

Portfolio value over time.

get_drawdown_curve

Drawdown (underwater) curve over time.

get_monthly_returns

Per-month return breakdown.

get_graph

Graph overview of the strategy's trades (nodes / edges).

🛡️

Robustness & optimization

3 tools
walk_forward

Out-of-sample cross-validation. Returns per-window IS/OOS and Walk-Forward Efficiency — the overfitting test.

monte_carlo

Bootstrap the trades to estimate risk of ruin and return/drawdown percentiles. Luck vs. edge.

optimize_exits

Sweep TP/SL/trailing ranges at fixed entries to find the best exit parameters.

📁

Reports — curate hundreds of runs

6 tools
list_reports

The curated list of every archived backtest and RL model the user sees.

save_report

Archive a run explicitly with a name and metrics.

open_report

Re-run a report's strategy and show it on the dashboard.

pin_report

Keep a winner (protected from auto-cleanup) or rename it.

delete_report

Drop a losing run.

clear_unpinned_reports

Bulk-delete everything that isn't pinned.

🧠

Reinforcement learning

4 tools
rl_train

Train a DQN trading agent (pure Rust) with a live reward curve and IS/OOS evaluation; the model is saved as a report.

rl_predict

Ask a trained model for the current signal (long / short / flat) on the latest window — or an inline bar batch.

rl_train_multi

Train a portfolio agent across up to 4 symbols at once (joint action space).

rl_predict_multi

Get a per-symbol signal from a saved portfolio model.

🎛️

App control

2 tools
get_ai_instructions

The system prompt + the current dataset's columns — the agent calls this first to orient.

execute_ui_action

Steer the app the user is watching: switch tabs, refresh, export, scroll to a section.

You never call these directly — your agent does. You just describe what you want in plain language, and watch the results render live in the app.