@romaco/mcp

MCP server for romaco-charts. Control your trading chart from Claude, Cursor, or any MCP-compatible AI agent.

npx @romaco/mcp

Philosophy

Romaco MCP is compression-first. Tools return features and decisions, not raw OHLCV. Most default tool payloads stay under 2 KB; romaco_analyze_market is the deliberate exception at 2.4–4.1 KB across the eight recorded 400-bar fixtures. Large payloads such as snapshots, visible-candle arrays, and all-bar indicator series are gated behind acknowledgeHighTokenCost: true. Full raw chart-state export is disabled until an explicitly authorized host contract exists.

The rule: code computes; the agent queries and interprets. RSI, levels, and last price come from structured tool output. This reduces unsupported numeric claims; it cannot prevent an agent from misquoting evidence or reaching a wrong thesis. Grounding data is not grounding the conclusion. Compression also preserves more context than returning a large raw OHLCV dump by default.

Cost table

Related MCP server: tradingview-mcp

30-second start

# 1. Install
npm install -g @romaco/mcp

# 2. Wire it into Claude Code (run from your project root)
cat > .mcp.json <<'EOF'
{ "mcpServers": { "romaco": { "command": "romaco-mcp" } } }
EOF

Then start Claude Code and prompt:

Use romaco_setup_chart to analyze AAPL daily with the trend_analysis preset.

That's it. The MCP server fetches yfinance data (with disk cache + cookie/crumb handshake), runs full technical analysis, and returns a compressed MarketSummary. No API key required; the cache reduces repeated upstream requests, but Yahoo can still rate-limit traffic and the server reports that failure explicitly.

Live chart control (optional)

To let Claude drive a real chart — drawings, indicators, alerts visible in your browser — mount <McpBridge /> next to your TradingTerminal:

import { TradingTerminal, McpBridge, type TradingTerminalRef } from 'romaco-charts/react';
import { useRef } from 'react';

function App() {
  const ref = useRef<TradingTerminalRef | null>(null);
  const bridgeToken = getBridgeTokenFromRuntime();
  return (
    <>
      <TradingTerminal ref={ref} data={candles} symbol="AAPL" />
      <McpBridge chartRef={ref} security={{ mode: 'paired', token: bridgeToken }} />
    </>
  );
}

See examples/pro-volatility-scanner for a complete setup. With <McpBridge /> mounted, the chart-bridge tools (add_indicator, add_drawing, add_alert, capture_snapshot, …) become available. Atomic thesis/pattern annotation additionally requires the host action documented in Chart bridge compatibility; older hosts fail closed instead of falling back to sequential drawing writes. romaco_annotate also uses a two-step, one-time confirmation challenge: its first call performs zero writes and returns APPROVAL_REQUIRED; retry only after the user approves the exact analysisId.

Data cache

Every successful yfinance fetch is cached to ~/.romaco/cache/ with per-timeframe TTL (1m→1min, 1h→1h, 1d→24h, …). A grid of N widgets requesting the same symbol triggers one upstream fetch. Wipe the cache anytime with rm -rf ~/.romaco/cache, or override the location with ROMACO_CACHE_DIR=/tmp/cache.

What it does

Exposes 20+ MCP tools in two categories:

Headless tools — work without a browser, load data and run analysis server-side:

• romaco_setup_chart — one-command setup: load + preset + analyze (recommended first call)

• romaco_load_candles — fetch OHLCV from Yahoo Finance (free) or pass your own array

• romaco_analyze_market — full technical analysis: trend, S/R levels, RSI/MACD/divergences, volatility, patterns

• romaco_thesis — computed bull/bear debate → verdict (long/short/stand_aside) + confidence + entry/stop/target setup; stands aside when R/R is poor (won't fake a signal)

• romaco_find_levels — support/resistance via K-means + Volume Profile (POC/VAH/VAL)

• romaco_detect_patterns — H&S, double top/bottom, triangles, flags

• romaco_calculate_position_size — pure-math risk-based position sizing with R/R + breakeven win-rate

• romaco_list_templates — catalog of drawing templates (trendline, fib, channels, …)

Chart-bridge tools — control a live Romaco chart in the browser:

• romaco_add_indicator — EMA, RSI, MACD, Bollinger, ATR, 29+ indicators

• romaco_add_drawing — agent-owned trendlines, Fibonacci, lines, channels, rectangles; omitted group defaults to romaco-mcp/manual

• romaco_add_alert — price alerts with direction (above/below/cross)

• romaco_clear_alerts — preview, approve, then remove exact alert IDs from one chart; never sends global clear

• romaco_capture_snapshot — PNG/JPEG base64 for vision LLMs

• romaco_open_paper_position — approval-gated simulated long/short with SL/TP and deterministic process-local idempotency

• romaco_get_chart_context — concise live chart state; raw chart export is disabled

• romaco_get_visible_candles — OHLCV in current viewport

• romaco_set_zoom / romaco_reset_view — zoom control

• romaco_clear_drawings — preview, approve, then remove only romaco-mcp/* groups; user drawings stay intact

• romaco_list_panes — enumerate main + subpanel panes (e.g. RSI subpanel id)

• romaco_get_indicator_values — read computed indicator series (by id or name)

• romaco_go_to_timestamp — scrub viewport to a given timestamp

• romaco_annotate — atomically draw one exact thesis after a scoped, one-time confirmation challenge

Install

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

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

Restart Claude Desktop.

Claude Code

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

Or add to your project's .mcp.json:

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

Cursor / other MCP clients

Any client that supports stdio MCP servers works. Point it at npx @romaco/mcp.

Usage

Headless analysis (no browser needed)

Load 500 candles of AAPL 1h from yfinance, then analyze the market.

Claude will call:

1. romaco_load_candles → fetches from Yahoo Finance

2. romaco_analyze_market → returns a compressed MarketSummary (2.4–4.1 KB on the recorded 400-bar fixture suite)

3. romaco_find_levels → S/R zones, POC, VAH, VAL

4. Reasons over the features → tells you what it sees

Live chart control (with browser)

Add <McpBridge /> to your chart app:

import { TradingTerminal, McpBridge } from 'romaco-charts/react';

function App() {
  const ref = useRef(null);
  const bridgeToken = getBridgeTokenFromRuntime();
  return (
    <>
      <TradingTerminal ref={ref} symbol="AAPL" timeframe="1h" datafeed={myDatafeed} />
      <McpBridge chartRef={ref} security={{ mode: 'paired', token: bridgeToken }} />
    </>
  );
}

Then from Claude:

Add EMA 20 and RSI 14 to the chart, draw a Fibonacci from the last swing low to swing high,
and capture a snapshot so I can see it.

romaco_annotate, romaco_clear_drawings, romaco_clear_alerts, and romaco_open_paper_position use two-step confirmation. First call returns a scoped token and performs zero chart writes. MCP rejects missing, expired, wrong-scope, and replayed tokens. Agent/client must send token only after explicit user approval. Token proves completion of protocol; by itself it cannot cryptographically prove human intent and never bypasses chart host's actionPolicy.

Drawing clear scope binds exact chart identity plus current Romaco-managed group plan. It never sends global clearDrawings; each reserved romaco-mcp/* group is replaced with empty desired state, leaving user and unrelated groups intact.

Alert clear scope binds exact chart identity plus current stable alert IDs. It never sends global clearAlerts; each approved alert is removed by ID with host identity policy. Alert-plan drift consumes the token and performs zero writes.

Paper positions are visual simulation only: no broker, real order, or money. Caller supplies stable idempotencyKey. Same completed key/payload returns same stored receipt without another chart write; changed payload gets IDEMPOTENCY_CONFLICT. Ambiguous bridge failure marks key indeterminate and blocks automatic retry until user inspects chart state, reconciles outcome, and uses a new key with fresh approval. Journal is memory-only for MCP process; restart/crash does not provide durable exactly-once execution.

Evaluation status

npm run eval:offline validates 28 declared scenarios and currently executes 24 deterministic MCP conformance workflows; 4 remain explicitly planned. Reported success rate uses executed runnable workflows as denominator and always reports planned work plus concrete blockers separately. Runnable failures or missing runnable workflow/grader implementations exit non-zero. Numeric agent-final-answer grading remains planned; current evidence-conformance checks structured tool evidence, not model prose.

Data sources

Bring your own key (BYOK) — Coming in v1.1: Alpaca, Polygon.io, FMP.

Configuration

Bridge security: paired v2 mutually authenticates server and browser with HMAC-SHA-256 before chart traffic. Token never crosses WebSocket. Generate one once, configure MCP process, inject same value into <McpBridge /> at runtime:

node -e "console.log(require('node:crypto').randomBytes(32).toString('base64url'))"

ROMACO_MCP_BRIDGE_AUTH=required \
ROMACO_MCP_BRIDGE_TOKEN="<generated-token>" \
ROMACO_MCP_ALLOWED_ORIGINS="https://myapp.com" \
npx @romaco/mcp

Never put token in source, public bundle, URL, logs, or browser storage. Use in-memory input for current page lifetime. auto selects paired v2 with valid token; no token keeps legacy v1 with warning. Malformed configured token disables only chart bridge. required never falls back to v1.

Legacy compatibility is unauthenticated and is not secure by default. Production deployments should use required plus paired McpBridge; enable legacy only as an explicit migration step.

Listener binds 127.0.0.1; exact origins add defense in depth. In paired mode, localhost pages still need pairing token. Browser-to-server frames above 8 MiB are rejected before JSON parsing, leaving more than 10x headroom over typical documented chart snapshots. Frames are not encrypted. Never expose listener remotely; use authenticated TLS gateway for non-loopback deployments.

# Custom port
npx @romaco/mcp --port 3200

# Or via env
ROMACO_MCP_PORT=3200 npx @romaco/mcp

Set the same port in <McpBridge port={3200} />.

Local execution and remote egress

Current tools compute analysis locally in src/compression. Setting ROMACO_TOKEN or ROMACO_API_URL does not authorize transmission of candles to a remote analysis service. When a token is present, romaco_thesis returns a structured REMOTE_EGRESS_DISABLED warning and uses the local artifact unless an already validated gateway artifact was injected through an authorized host path.

A future remote adapter must add explicit authorization, validated boundary schemas, transport security, trace propagation, and user-facing egress docs before it can become callable. No automatic network fallback exists today.

Example prompts

Institutional analysis:

"Load TSLA 4h from yfinance, analyze the market, find key levels, detect any chart patterns, then draw the Fibonacci retracement of the last swing on the chart and add an alert at the 0.618 level."

Conditional automation:

"Load AAPL 15m, run full analysis. If RSI shows bullish divergence and price is near a support level, open a paper long 100 shares with stop loss at the VAL of the Volume Profile."

Pattern scanner:

"Load SPY 1d and detect patterns. For any head & shoulders found, tell me the target price and invalidation level."

Requirements

• Node.js >= 20

• For chart-bridge tools: romaco-charts >= 1.0.0-beta.6 with <McpBridge /> in your app

Links

• romaco-charts npm

• Documentation

• romaco.io

• GitHub