traderkit

Risk-gate MCP server for AI-assisted trading. Enforces position caps, wash-sale rules, and forbidden-structure checks before any order hits your broker.

Built for Claude Code + SnapTrade. Works with any MCP client.

Install

npx -y traderkit

Or globally:

npm install -g traderkit

What it does

Seventeen MCP tools that sit between your AI assistant and your broker:

check_trade

The primary gate. Validates:

• Order size cap — rejects if notional_usd exceeds max_order_notional
• Concentration cap — rejects if post-trade single-name exposure exceeds max_single_name_pct
• Forbidden tools — blocks specific broker tools (e.g., margin tools)
• Forbidden leg shapes — blocks option structures (e.g., naked_put, naked_call)
• Wash-sale check — flags if same ticker was sold at a loss within ±30 days (same tax_entity)

Returns { pass: boolean, reasons: string[], warnings: string[] }.

check_wash_sale

Standalone wash-sale check. Pulls last 30 days of activity from a sibling snaptrade-mcp-ts server. Pools all accounts under the same tax_entity (e.g., all personal accounts share one wash-sale window; an LLC has its own).

Graceful degradation: if snaptrade-read is unavailable, returns flagged: false with a warning rather than blocking.

scan_tlh

Scans your positions for tax-loss harvesting opportunities. Filters to positions with unrealized loss above a threshold (default $500), then excludes any that would trigger a wash sale. Returns candidates sorted by loss size (largest first).

Requires positions data as input (from snaptrade_get_positions or equivalent).

check_concentration

Analyzes portfolio concentration against profile caps. Returns every position labeled:

• HEADROOM — well below cap
• AT-CAP — 75-90% of cap
• NEAR-CAP — 90-100% of cap
• OVER-CAP — exceeds cap

Also returns the HHI (Herfindahl-Hirschman Index) as a single-number diversification score, and the top 5 positions.

regime_gate

Checks if a trade should proceed under the current market regime:

Returns adjusted notional, preferred structures for the tier, and whether the proposed structure aligns.

propose_trade

End-to-end trade proposal builder. Takes a ticker, price, portfolio context, and regime tier. Produces a fully sized proposal with:

• Headroom-based sizing — (cap% - current%) × 0.5 × NAV × regime_multiplier
• Cap enforcement — capped at profile's max_order_notional
• Regime blocking — rejects BUY in DEFENSIVE/HALT
• Concentration check — rejects adds when already OVER-CAP
• Sizing trace — human-readable formula for audit

Optional: attach thesis_ref and signal_summary for proposal context.

track_tax

Computes running tax exposure from an array of realized trades:

• Separates STCG (<365 days) from LTCG (≥365 days)
• Computes reserves at configurable rates (defaults: STCG 35.8%, LTCG 18.8%)
• Tracks gains and losses separately per bucket
• Flags wash-sale-adjusted trades
• Returns per-trade breakdown sorted by date

trigger_check

Event detector for portfolio monitoring. Checks three conditions:

• NAV_MOVE — triggers on ±2% NAV change (configurable). CRITICAL at ±4%.
• REGIME_SHIFT — fires when regime tier changes. CRITICAL on deterioration, INFO on improvement.
• CONCENTRATION_BREACH — flags positions exceeding the cap. CRITICAL when 10pp+ over.

Returns events sorted by severity (CRITICAL → WARNING → INFO).

performance_metrics

Computes portfolio performance metrics from a daily returns series:

• Sharpe ratio — risk-adjusted return (sample variance, annualized)
• Sortino ratio — downside-only volatility version
• Max drawdown — largest peak-to-trough decline (fraction 0–1), with peak/trough indices
• Calmar ratio — annualized return / max drawdown
• Win rate — fraction of positive returns, plus average win and average loss

Requires minimum 20 observations by default (configurable via min_observations). Supports custom risk_free_rate (default 5%) and periods_per_year (default 252).

thesis_fit

Scores how well a trade aligns with active theses:

• IN_THESIS — ticker + structure match an active thesis
• PARTIAL — ticker matches but structure doesn't
• OFF_THESIS — ticker not in thesis, or thesis is closed
• NO_THESIS_REF — no thesis_ref provided

Supports score_fit (single) and batch_score (portfolio-wide). Batch returns a summary with counts per score tier.

trading_calendar

NYSE trading calendar with 10 observed holidays + Good Friday. Five actions:

• is_trading_day — check if a date is a trading day
• next_trading_day / prev_trading_day — find adjacent trading days
• last_trading_day_of_month — useful for monthly rolls and reviews
• trading_days_between — count trading days in a range

session_write

Formats session document sections as markdown. Four actions:

• format_executed — markdown table of executed trades
• format_deferred — bullet list with deferred tags
• format_no_trade — bullet list of rejected tickers with reasons
• format_session_index_row — 11-column index table row for session tracking

broker_route

Classifies broker routing for order dispatch:

• SNAPTRADE — Fidelity, E-Trade, Robinhood, Schwab, IBKR
• TRADESTATION — TradeStation
• MANUAL — Ally, Morgan Stanley, unknown brokers
• DEFERRED — any trade with deferred tags (overrides broker classification)

Setup

1. Create profiles

Profiles live in ~/.traderkit/profiles/ as markdown files with YAML frontmatter:

---
name: personal
broker: snaptrade
account_id: xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
tax_entity: personal
caps:
  max_order_notional: 10000
  max_single_name_pct: 25
  forbidden_tools: []
  forbidden_leg_shapes: [naked_put, naked_call]
---

Get your account_id from snaptrade_list_accounts.

2. Register in Claude Code

Add to .claude/settings.json:

{
  "mcpServers": {
    "traderkit": { "command": "npx", "args": ["-y", "traderkit"] }
  }
}

3. Wire the PreToolUse hook

The companion PreToolUse hook calls check_trade automatically before any destructive broker tool fires. Add to .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "mcp__snaptrade-trade__equity_force_place|mcp__snaptrade-trade__mleg_place",
        "command": "node ~/.traderkit/scripts/pre-tool-use.js"
      }
    ]
  }
}

4. Verify

traderkit  # starts MCP server on stdio

In Claude Code: ask "list profiles" to confirm the server is connected.

Environment variables

How it works

Claude Code ──PreToolUse hook──► traderkit MCP
                                    │
                                    ├─ caps check (profile YAML)
                                    ├─ wash-sale check (snaptrade-mcp-ts)
                                    │
                                    ▼
                              pass/block decision

• Fail-closed by default. If traderkit can't evaluate a trade, it blocks.
• Credential redaction. All tool responses are scrubbed — any env secret substring (8+ chars) is replaced with <REDACTED>.
• Tax-entity pooling. Wash-sale checks span all accounts with the same tax_entity. Personal brokerage + IRA = one pool. LLC = separate pool.

Supported brokers

Works with any broker connected via SnapTrade:

• Fidelity (read + write)
• E-Trade (read + write)
• IBKR (read + write)
• Schwab (read + write)
• Robinhood (read-only)
• TradeStation (via separate TradeStation MCP)

Full setup

For the complete trading terminal setup (vault templates, profiles, scripts, docs), see the traderkit repo.

License

MIT