# UnityBridge-MCP

A normalisation + MCP layer for trade data. Heterogeneous messages — **FIX**,
**venue JSON**, **broker CSV** — land in different shapes; UnityBridge maps them
all to one normalised record and exposes that over MCP, so an LLM or agent can
query it in plain English without touching the raw plumbing.

This mirrors the job Quod's Unity layer does: a normalised, real-time data layer
that any downstream consumer (including AI) can read against one schema.

```
   FIX 4.4 ─┐
   JSON     ├─► normalizer ─► NormalizedOrder ─► OrderStore ─► MCP tools
   CSV      ┘   (detect +        (one schema)     (query +      query_orders
                 parse +                           aggregate)    aggregate_latency
                 validate)                                       ingest_stats
```

## What's real here

- `parsers/fix.py`, `parsers/venue_json.py`, `parsers/broker_csv.py` — three real
  parsers with correct field maps (FIX tags 11/55/54/38/44/30/39, venue key
  aliases, positional CSV). Each maps to the same `NormalizedOrder`.
- `normalizer.py` — format detection, dispatch, pydantic validation, latency
  enrichment, and **error capture** (malformed messages are recorded, not
  dropped silently — the success rate is a real metric).
- `store/order_store.py` — filterable query + aggregates (p50 latency by venue,
  reject rate, counts by status), computed, not hard-coded.

## Quick start

```bash
make install
make replay N=500          # replay a synthetic feed, print normalisation stats
make mcp                   # serve query tools over MCP
make test
docker compose up
```

## Status

Reference skeleton. Parsers, normaliser and store are real and run offline on a
synthetic multi-format feed. `# INTEGRATION:` markers show where the live Unity /
Kafka bus replaces the synthetic generator.