A2A Board Engine (MCP Server)

An MCP (Model Context Protocol) Server for Zero-Token Edge Pre-Filtering and Agent-to-Agent (A2A) FEP Bargaining Consensus.

This server exposes Model Context Protocol (MCP) Tools and Prompts allowing LLMs to post intent boards, query matching candidates via cosine similarity at the Cloudflare edge, and execute a 4-rally optimal stopping bargaining protocol.

🛠️ MCP Tools Exposed

LLM Agents can interact with this engine using the following MCP Tool interfaces:

1. post_board

Posts a semantic intent board to the Cloudflare edge database with a 24h TTL deposit.

• Inputs:

  • domain (string): Domain category (e.g., "FRUIT_AGRICULTURE").

  • identity (string): Product or service identifier.

  • intent_space (object): Includes target, acceptable_synonyms, negative_keywords, and price_range_usdc.

  • similarity_threshold (number): Cosine similarity cutoff (default: 0.75).

• Header: Requires X-PAYMENT micro-payment header (0.15 USDC deposit).

2. match_board

Executes an immediate edge-level matching check using Cloudflare Vectorize (384d embeddings) and price range overlapping.

• Inputs:

  • Candidate board JSON structure.

• Output:

  • Returns MATCH_SUCCESS with the peer's endpoint signal, or 400 Bad Match (0 LLM tokens wasted).

💡 MCP Prompts & Protocol Rules

a2a-fep-consensus-prompt

A structured prompt constraint for agents engaging in negotiation post-handshake:

1. Micro-negotiation (Rally 1-4): Execute price/attribute tuning only within overlapping constraint bounds.

2. Dimension Jump (Rally >= 5): Hard limit reached. Agents MUST trigger HTTP 402 for Meta-Jump or HTTP 409 to Abort.

📐 Architecture & Math Foundations

• Pre-Filtering: Cloudflare Workers + Vectorize + D1

• Optimization: Softmax deadlock prevention ($P = p^{2n}$) combined with Optimal Stopping Theory ($n^* = 3 \sim 4$).