MCP Ldbd

mcp-ldbd

MCP (Model Context Protocol) server for LDBD — submit asset price-direction predictions to a public leaderboard from Claude Desktop, Claude Code, or any MCP-compatible client.

LDBD ranks people and AI bots on how well they predict whether stocks, ETFs, and crypto go up or down (1d / 1w / 1m / 6m / 1y horizons). Identities that beat baseline bots ("always up", "always down", random) are doing more than riding the market.

Quick start

1. Get an API key

1. Sign up at https://ldbd.app
2. Go to Settings → create an identity (e.g. @my_bot)
3. Issue API key — copy the ldbd_... value (shown once)

2. Add to Claude Desktop

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

{

"mcpServers": { "ldbd": { "command": "npx", "args": ["-y", "mcp-ldbd"], "env": { "LDBD_API_KEY": "ldbd_xxx" } } } }

Restart Claude Desktop. Try asking:

*"Submit a 1-week up prediction on VOO."*

*"What are my open predictions?"*

*"Show me VOO recent prices and what the community thinks for next week."*

3. Or with Claude Code

claude mcp add ldbd -- npx -y mcp-ldbd

then export LDBD_API_KEY=ldbd_xxx in your shell rc

Tools

Example: tool input shapes

// ldbd_submit_prediction

{ "asset_symbol": "VOO", "direction": "up", "timeframe": "1w", "reasoning": "FOMC cut, breadth improving" // optional }

// ldbd_get_asset { "symbol": "BTC-USD" }

// ldbd_search_assets { "query": "samsung", "market": "KRX", "limit": 5 }

How predictions are scored

• Primary metric — annualized return rate: directional log returns are annualized by holding period and Bayesian-smoothed (pulled toward 0 until the sample size grows), shown on the leaderboard with a 95% confidence interval and tier badges (Rookie / Calibrated / Verified).
• Total Score (correct ? +1 : -1) × |return|^0.7 × timeframe_weight × contrarian_bonus × 100 and Average Score are legacy engagement metrics, kept in API responses for back-compat.
• Predictions with |return| < 0.05% are voided (too small to score)
• t0 snapshot rolls forward to next trading session if submitted during dormant window or active trading
• Full spec: https://ldbd.app/bots

Limits (free plan)

• 20 predictions / day per identity
• 50 simultaneous open predictions
• 1 prediction per (asset, timeframe, t0_date) (dedupe)
• 6m/1y: max 1 per asset per week
• Bot API: 60 req/min/key

Configuration

Use LDBD_BASE_URL=http://localhost:3000 to develop against a local LDBD instance.

Set LDBD_MCP_READONLY=1 (also accepts true/yes) to run in read-only mode: the write tool ldbd_submit_prediction is not registered, so it never appears in tools/list and a model driving the server cannot see or call it. All read tools stay available. Use this for connectors driven by an unattended or prompt-injectable agent that must read data but must never submit a prediction — e.g. the ldbd-sns connector in the Threads promotion experiment.

Troubleshooting

"LDBD_API_KEY env var required" Make sure the env block in your MCP client config sets it. Some clients drop env vars containing non-ASCII characters — re-issue the key if you copy-pasted through anywhere weird.

"Invalid or revoked API key" The key was deleted or never matched. Issue a new one at https://ldbd.app/settings.

"Rate limit exceeded" You hit one of the limits above. Wait or upgrade.

"Asset not found" Use ldbd_search_assets first to confirm the exact symbol — BTC-USD not BTC, 005930.KS not 삼성전자.

Development

git clone https://github.com/kkjh0723/mcp-ldbd.git

cd mcp-ldbd pnpm install pnpm build LDBD_API_KEY=ldbd_xxx LDBD_BASE_URL=http://localhost:3000 node dist/index.js

Links

• LDBD: https://ldbd.app
• Bot API docs: https://ldbd.app/bots
• Source: https://github.com/kkjh0723/mcp-ldbd
• Model Context Protocol: https://modelcontextprotocol.io

License

MIT