io.github.doazvjettu/leakguard-mcp

leakguard-mcp

Squawk for backtests. A local-first MCP server that static-analyzes agent-generated Python code and flags lookahead bias & data leakage before the backtest runs.

Works for any time-series ML code — quant trading (crypto, equities, forex, futures), demand forecasting, energy, weather, IoT sensors — wherever a wrong .shift() or a global normalization silently poisons your results.

• Pure source analysis (libcst) — your code never leaves the machine, no API calls.
• Heuristic, not a proof: severity tiers (🔴 error / 🟡 warning), like an aviation squawk.
• Framework-agnostic: pandas / numpy / polars, any stack.
• MCP-native (stdio): works directly inside Claude Code, Cursor, and any MCP-compatible agent.
• All 10 rules ship free — no license, no tiers, no phone-home.

Why this exists

AI agents (Claude Code, Cursor) write feature engineering and strategy code faster than humans can review it. But they introduce lookahead bias at scale — subtle time-boundary errors that backtest perfectly and fail catastrophically in live trading or production:

# Agent writes this — looks fine, is catastrophically wrong
df['momentum'] = df['close'].shift(-5)   # uses FUTURE prices as a feature
df['vol_norm'] = (df['close'] - df['close'].mean()) / df['close'].std()  # leaks future mean

leakguard catches these in the same agent loop — before the backtest runs:

LG001 🔴 line 2: Future shift used as feature — shift(-5) references 5 bars ahead.
  Fix: df['momentum'] = df['close'].shift(5)   (lag, not lead)

LG003 🔴 line 3: Global-fit normalization — mean/std computed over the full series
  before any train/test split, leaking future statistics into the past.
  Fix: df['vol_norm'] = (df['close'] - df['close'].expanding().mean()) / df['close'].expanding().std()

The agent reads the finding + fix snippet and self-corrects in one turn. No human review needed.

Install

Requirements

• Python 3.11+
• uv (recommended) or pip

From PyPI

pip install leakguard-mcp

From source

git clone https://github.com/doazvjettu/leakguard-mcp
cd leakguard-mcp
uv sync

Setup with Claude Code

Add to your MCP config (~/.claude/claude_desktop_config.json or .claude/settings.json in your project):

{
  "mcpServers": {
    "leakguard": {
      "command": "python",
      "args": ["-m", "leakguard.server"]
    }
  }
}

Or if installed via uv:

{
  "mcpServers": {
    "leakguard": {
      "command": "uv",
      "args": ["run", "python", "-m", "leakguard.server"]
    }
  }
}

Restart Claude Code. leakguard's tools are now available to the agent.

Setup with Cursor

Add the same block under mcpServers in your Cursor MCP settings file.

MCP Tools

CLI

The same scanner is available as a CLI — handy for a pre-commit hook or CI step (exits non-zero when leakage is found):

uv run leakguard path/to/strategy.py
# or, installed:  leakguard path/to/strategy.py

It prints each finding with its severity, line/col, and a concrete fix snippet — the same output shown in the demo above.

Rules

All 10 rules active, no tiers:

Each finding includes a concrete fix snippet so the calling agent can self-correct immediately.

Benchmark

Measured on two labeled corpora, 49 snippets total. Reproduce with uv run python -m benchmark.run.

Honesty note: the trading corpus was written by the tool's author — treat its numbers as regression fixtures, not independent validation. The general-ML corpus is one arm's length removed in domain (author-composed reproductions of widely documented leakage anti-patterns, not a downloaded public dataset). The corpus deliberately includes adversarial snippets the scanner is known to miss; they are counted against it.

Trading corpus — 39 snippets (23 leaky, 16 clean + hard negatives):

General-ML corpus — 10 snippets (LG003/LG004/LG010): Precision 88%, Recall 100% (TP 7 / FP 1 / FN 0).

Combined: Precision 90.4%, Recall 100% (TP 47 / FP 5 / FN 0).

Recall is 100% on this corpus — every adversarial miss exposed has since been fixed (constant propagation, hand-rolled normalization, cv=<int>, drop-based selection). Leak shapes not yet in the corpus are still missed — see Known Limitations below.

Known false positives (clean code that gets flagged)

• LG001: a forward-return label built with a negative shift and used only as y — pure AST cannot distinguish a target column from a feature.
• LG003: fit inside a helper defined above the split call site — line-order heuristic confuses definition order with execution order.
• LG004: shuffled splits on genuinely non-temporal data — no datetime-index inference yet.
• LG009: resampling for reporting/plotting rather than features — intent is invisible to static analysis.

Known false negatives (leak shapes not yet covered)

• LG004: cross_val_score(...) with cv omitted (defaults to KFold).
• LG005: taint through df.loc[:, 'col'] = ... or df.assign(col=...).
• All rules: values flowing through function calls, dicts, or non-constant variables — no cross-function dataflow.

These sets are pinned in tests/test_benchmark.py: any new miss or silent fix fails the suite until docs and corpus are updated to match.

Develop

uv sync --extra dev
uv run pytest                              # 96 tests
uv run python -m leakguard.server          # stdio MCP server
uv run leakguard demo/strategy_leaky.py    # CLI on the demo file
uv run python -m benchmark.run             # precision/recall tables + FP/FN lists

The scanner core lives in leakguard/core/ (pure, no MCP imports); server.py and cli.py are thin wrappers. Each rule has a fixture pair under tests/fixtures/.

Limitations (v1)

• Heuristic static analysis — catches ~90% of common patterns, not 100%.
• Single-file only — no cross-module taint tracking.
• Python only (pandas / numpy / polars).
• No runtime execution — cannot catch patterns that only emerge at runtime.

Contributions welcome: new corpus snippets (especially real bugs you've hit) strengthen the benchmark more than new rules do.