A confidence score on a double top says how well the shape fits, not whether the shape has ever mattered; here the lift against baseline answers the second question and often comes back as noise. The delta call is the other design decision worth having — monitoring twenty tickers twice a day costs near nothing in context.
A market-data service for agents covering US stocks, ETFs and crypto spot. One call with a ticker and a timeframe returns an interpreted state report — compact candles, detected chart and candlestick patterns, support and resistance, regime, RSI and EMA — instead of raw OHLCV for the model to chew through. Every detected pattern carries its backtested hit rate and its lift against the same market with no pattern present.
- A market-state brief already interpreted: for BTC/USDT at 4h over 120 candles, about 1,323 tokens of analysis in place of roughly 3,260 tokens of numbers the model still has to read
- Evidence attached to each pattern — sample count, hit rate, a 95% confidence interval, and the lift over baseline, with many patterns honestly returning indistinguishable-from-baseline so an agent can filter on whether a signal informs anything
- A delta call that returns only what moved since the last brief — trend flips, new patterns, RSI state changes — and says changed: false when nothing did, which is what makes polling a watchlist affordable
- A multi-timeframe read in one call with the alignment spelled out, so 1h up / 4h up / 1d down is visible as a conflict rather than inferred from three separate answers
- Historical analogs from the same series: earlier windows whose shape resembles now, with the full distribution of what followed — win rate, median, min, max and n over a fixed forward horizon
- A screener over liquid US large caps, core and sector ETFs and major crypto pairs, filtered by asset class, regime, pattern and minimum base rate, ranked by base rate with its confidence interval
- A capability call that states the covered asset classes, data sources and delays, timeframes, endpoints, per-call prices and tier limits, and returns no market data at all
- Timeframes from 1m through 1w, with US equities split- and dividend-adjusted and crypto spot in realtime
Nothing installed: the remote server at https://patternfetch.com/mcp speaks streamable HTTP, and initialize and tools/list are free — only a tool call needs auth. Add the URL in Claude Code, Claude Desktop, Cursor or claude.ai and authorize once, which mints a free-tier key, or send a pf_ key as a bearer token. For stdio-only setups the npm package ships patternfetch-mcp, a zero-dependency bridge that takes PATTERNFETCH_API_KEY and falls back to an embedded snapshot for tools/list when the remote is unreachable. Calls are priced individually and paid from credit — $3 on signup, at least $0.50 of it usable without a card. The output is impersonal market data, not advice: base rates are gross directional frequencies with no stops, fees or slippage modelled.
One command — npx -y patternfetch-mcp