The response is built for a model to act on: one signal, one confidence number, and the specific book and tape measurements that produced them, so an agent can check the reasoning rather than take the label on trust. Symbols without enough data come back in a data-gathering state instead of being scored, which is the right behaviour when something is about to act on the answer. The cost is a RapidAPI plan.
An MCP server over the Horus Flow Intelligence API. Rather than returning candles for a model to interpret, it returns a decision-shaped object: a signal, a confidence value, a market state and a risk level, with the order-book and tape measurements that produced them attached. The feeds behind it are the Binance L1/L2 book and Alpaca equities.
- A structured read per symbol — `signal`, `confidence`, `market_state` and `risk`, with a short description of what the engine saw
- The measurements behind the verdict: `bid_ask_ratio` and `buy_sell_ratio` for book and taker imbalance, `delta_5s` for recent flow, `delta_accel` for how fast it is changing, plus `whale_activity` and `large_sell_count`
- Behavioural flags rather than raw numbers alone — `SPOOFING_DETECTED` with the side of the wall, and `GLOBAL_LIQUIDITY_EVENT`
- A broader climate read alongside the symbol-level one — `wiseman_climate` carries market mode, health and its own confidence
- The same data over plain HTTP for non-MCP callers, at `https://flow.horustek.pro/v1/flow/crypto/BTCUSDT` with an `X-API-Key` header
Python 3.12+ and a RapidAPI key for Horus Flow Intelligence, exported as `RAPIDAPI_KEY` — the server needs it to fetch live data. `pip install mcp httpx`, then `python horus_mcp_public.py --transport sse --port 8011` for SSE. The published package is `horus-flow-mcp` (1.0.6 in pyproject), which `server.json` declares over stdio with `RAPIDAPI_KEY` marked required and secret. Data access is a paid RapidAPI plan, separate from the server itself.
One command plus a key — uvx horus-flow-mcp, then supply credentials