@trendcraft/mcp

Model Context Protocol server for TrendCraft — exposes the indicator manifest (96+ entries with whenToUse / pitfalls / synergy / marketRegime / timeframe / paramHints) plus calc and signal dispatchers to LLM clients like Claude Desktop and Cursor.

Positioning: this is an indicator knowledge server, not a backtest wrapper. The differentiator is the structured per-indicator metadata that lets an LLM decide which tool to use, paired with a token-aware calc/signal API designed for agentic workflows.

Install

npm install -g @trendcraft/mcp
# or run on demand
npx @trendcraft/mcp

Configuring an MCP client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "trendcraft": {
      "command": "npx",
      "args": ["-y", "@trendcraft/mcp"]
    }
  }
}

Restart Claude Desktop.

Claude Code

claude mcp add trendcraft -- npx -y @trendcraft/mcp

Or edit ~/.claude.json directly under the mcpServers key with the same config as Claude Desktop.

Cursor

Cursor Settings → MCP → Add new MCP Server:

{
  "mcpServers": {
    "trendcraft": {
      "command": "npx",
      "args": ["-y", "@trendcraft/mcp"]
    }
  }
}

Verify

After connecting, the seven tools below should appear in the client's tool picker. To smoke-test the binary directly without an MCP client:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
  | npx -y @trendcraft/mcp

Tools

| Tool | Purpose | |---|---| | list_indicators | Discover indicator kinds. Returns kind, displayName, oneLiner, category, and calcSupported per entry. Optional filters: category, regime, timeframe, calcSupported. | | get_indicator_manifest | Full IndicatorManifest for one kind — whenToUse, pitfalls, synergy, paramHints, etc. Throws UNKNOWN_KIND if the manifest entry is missing. | | suggest_indicators_for_regime | Indicators well-suited to trending / ranging / volatile / low-volatility. | | format_manifest_markdown | Render one indicator's manifest as Markdown — useful for agent prompt embedding. | | calc_indicator | Compute one indicator on caller-supplied OHLCV candles. ~60 kinds have safe-calc wrappers. lastN (default 200) trims response size. | | list_signals | Discover signal kinds supported by detect_signal. Returns kind, shape (series or events), oneLiner, and paramsHint. Optional shape filter. | | detect_signal | Detect a trading signal — crossovers, MA alignment, candlestick patterns, divergences, squeeze, and volume signals. Returns { output, firedAt, ... } where firedAt is a sparse list of trigger times — designed for cheap screening ("did the signal fire in the last N bars?"). |

Designed for screening

The output shape is tuned for two LLM-driven workflows: single-symbol analysis (call a few indicators + signals on one symbol's candles) and multi-symbol screening (loop over many symbols, ask one yes/no per signal). For screening, detect_signal returns a firedAt: number[] summary so the caller doesn't have to scan the full series — typically 100x cheaper on tokens than reading the boolean output array.

// detect_signal output for a screening question
{
  "kind": "goldenCross",
  "shape": "series",
  "totalLength": 80,
  "count": 80,
  "truncated": false,
  "output": [/* full Series<boolean> here */],
  "firedAt": []   // ← scan this. empty = no fire in window.
}

Recipes

See EXAMPLES.md for end-to-end patterns: regime-driven discovery, multi-symbol screening, working with series-vs-events signals, indicator synergy, token-budget tuning, and error recovery.

Example workflow

A typical agent flow against a single symbol's daily candles:

// 1. Discover what to use for the regime you suspect.
suggest_indicators_for_regime({ regime: "trending" })

// 2. Look up paramHints / pitfalls before computing.
get_indicator_manifest({ kind: "rsi" })

// 3. Compute on the symbol's candles. lastN keeps the response compact.
calc_indicator({
  kind: "rsi",
  candles: /* { time, open, high, low, close, volume? }[] */,
  params: { period: 14 },
  lastN: 5
})

// 4. Discover signal kinds and their parameter shape.
list_signals({ shape: "series" })

// 5. Check whether a screening signal fired recently.
detect_signal({
  kind: "goldenCross",
  candles,
  params: { short: 5, long: 25 },
  lastN: 10
})
// → { firedAt: [...timestamps], output: [...] }

Errors

All tool errors return isError: true with the message body in the form <CODE>: <message>. Codes:

| Code | Meaning | Surface | |---|---|---| | INVALID_INPUT | Empty / malformed candles array | calc, detect_signal | | INVALID_PARAMETER | Missing or invalid params (negative period, missing required params object, etc.) | calc, detect_signal | | INSUFFICIENT_DATA | Candles count below the indicator/signal's minimum required window | calc, detect_signal | | INDICATOR_ERROR | Underlying indicator raised an unclassified runtime error | calc | | SIGNAL_ERROR | Underlying signal raised an unclassified runtime error | detect_signal | | UNSUPPORTED_KIND | Manifest entry exists but no calc wrapper is available | calc | | UNSUPPORTED_SIGNAL | Kind is not registered with detect_signal | detect_signal | | UNKNOWN_KIND | No manifest entry exists for the given kind | manifest tools | | INTERNAL_ERROR | Wrapper returned an unexpected envelope (should not occur) | calc, detect_signal |

UNSUPPORTED_KIND (calc) and UNKNOWN_KIND (manifest) are intentionally distinct: the former means the kind is documented in the manifest but the safe-calc wrapper hasn't been written yet; the latter means the kind doesn't exist at all. Use list_indicators({ calcSupported: true }) to enumerate the calc-ready set.

Data sourcing

calc_indicator and detect_signal both expect candles supplied by the caller. The server has no built-in data connector — pair it with another MCP server (Yahoo / Alpaca / your broker) or have the agent paste candles from a local file. This is a deliberate scope decision: data fetching, auth, and rate-limiting belong in dedicated tools.

Development

pnpm install --frozen-lockfile
pnpm --filter trendcraft build      # core must build first
pnpm --filter @trendcraft/mcp build
pnpm --filter @trendcraft/mcp test

Quick stdio smoke test:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
  | node packages/mcp/dist/bin/trendcraft-mcp.js

Disclaimer

@trendcraft/mcp and the underlying trendcraft library provide technical analysis primitives for informational and educational purposes only. Outputs — including indicator values, signal triggers, and manifest interpretive notes — are not investment advice and do not constitute a recommendation to buy, sell, or hold any financial instrument. You are solely responsible for any trading decisions made using this software.

License

MIT