Multi venue MCP

MaigaXBT’s AI agent leverages a multi-venue MCP Server

1. Connector Layer: CEX & AMM Adapters

Each venue (Binance, OKX, Uniswap, Hyperliquid, etc.) is fronted by a Connector microservice that implements a unified interface

• CEX adapters talk REST/WebSocket APIs, handle rate-limits, and parse order-book, trade, and funding-rate endpoints.

• AMM adapters poll on-chain subgraphs or use direct RPC calls to fetch get_pool_state (reserves, fees, token balances) and assemble “virtual order books” by inverting the constant-product curve.

All connectors register themselves with the MCP Server’s service registry, advertising capabilities (supports_order_book, supports_swap, etc.).

2. MCP Server: Context Aggregation & Normalization

The MCP Server performs:

1. Fan-Out Fetch

   • In parallel, calls each Connector’s get_order_book, get_trades, and, for AMMs, get_pool_state.

2. Normalization

   • Maps CEX bids/asks and AMM “swap curve samples” into a unified OrderBook object

   • Embeds pool reserves & implied price tiers under amm_markets.

3. Delta Tracking

   • If since is provided, only deltas (new/updated price levels, trades, pool changes) are included to minimize payload.

4. Merge & Return

   • Aggregates all venue data into one MCP Context Packet

3. Agent Inference: Signal Processing & Analysis

Inside a TEE-protected enclave, the MaigaXBT agent:

1. Ingests the MCP Context Packet (JSON) and:

   • Converts order books into time-series arrays.

   • Transforms AMM reserves into marginal price curves.

   • Extracts funding-rate trends and volume aggregates.

2. Feature Engineering

   • Computes cross-venue liquidity imbalances, price spreads, funding deviations, and on-chain flow ratios.

   • Applies sliding-window volatility measures and social-sentiment embeddings (if multimodal inputs are provided).

3. Model Inference

   • Passes engineered features into a neural network (e.g., Llama-based or custom MLP) to output

4. Risk Check (in-enclave)

   • Verifies aggregate exposure limits (e.g., max 10 BTC open across venues).

   • Applies position-size caps based on current volatilities.

4. Execution: Automated Trade & Swap Calls

The agent then posts its orders to the MCP Server:

1. Smart Routing

   • For CEX orders: may further split a large 2.5 BTC buy across Binance and OKX based on real-time depth.

   • For AMM swaps: calculates optimal path and slippage tolerances.

2. Atomic Execution

   • Uses either native OCO/ICEBERG features on CEXs or bundling via Flashbots (for AMM swaps) to ensure either all legs succeed or are rolled back.

3. Fill Reporting

   • Streams back a combined FillReport with per-venue fills, average fill prices, and gas/fee breakdowns.

5. Reporting & Post-Trade Analysis

All context packets, signal decisions, and fill reports are logged to a time-series store. Downstream “Reporting Services” can:

• Generate PnL Dashboards

  • Query aggregate fills vs. average execution price across CEX+AMM.

• Collateral & Fee Reports

  • Break down gas spent on AMM swaps vs. CEX taker fees.

• Performance Analytics

  • Compare model confidence vs. actual profit, refine model hyperparameters.

6. End-to-End Automation Loop

1. Loop Scheduling

   • A cron-style scheduler inside the agent enclave triggers every minute (or on event hooks) to fetch a new MCP context.

2. Continuous Learning

   • Execution and outcome data feed back into an on-chain “Proof-of-Trading” smart contract; later batch-pulled for model retraining.

3. Failover & Alerts

   • If a venue connector fails, the MCP Server signals the agent, which can either pause or reroute orders automatically.

4. Audit & Compliance

   • All requests/responses are signed inside the enclave; an off-chain auditor can verify the chain of custody before funds are unlocked in PoT.