System Architecture

ETMCP's system architecture is designed to support scalable, secure, and modular agent-based compute across decentralized networks. Inspired by the DeMCP structure, our architecture is composed of the following layers:

1. Client Layer

• Users & Agents: End-users and AI agents serve as primary input/output entities.

• Client Applications: Interfaces such as desktop apps or Web3 dApps (e.g., Copilot Studio, Claude Desktop, etc.) allow user interaction.

2. Language Model Layer

• LLM (Large Language Model): Acts as the cognitive engine for understanding prompts and deciding which tools or agents to invoke.

• Interacts with MCP Clients for execution tasks.

3. MCP Client Layer

• MCP Client: The primary coordinator that interacts with registries, data stores, and TEE nodes.

• It receives prompts/tool calls and securely delegates workloads.

4. Registry & Verification

• Centralized MCP Registry: A discovery layer where verified metadata about tools and agents are maintained.

• Blockchain Registry: On-chain ledger for registry of agents, smart contracts, and proofs.

• Attestation Verifier: Validates execution integrity using zk/TEE attestation methods.

5. Execution & Secure Environment

• TEE MCP Server: Executes tasks in a secure enclave, includes:

  • Tools Execution Engine

  • I/O Encryption Processor

  • Security Policy Enforcer

• Trust Execution Environment: Ensures workloads are run in isolated, tamper-proof environments.

6. Storage & Libraries

• User Data Storage: Handles temporary or persistent state of agents and data context.

• Function Library / APIs / Resources: A pool of callable external functions or models that agents may invoke.

7. Flow Summary

1. A user prompt is submitted through a client app.

2. The LLM interprets the prompt and calls an agent/tool.

3. MCP Client verifies registry data and delegates execution.

4. Secure execution takes place via TEE MCP Server.

5. Result is verified, returned to user, and optionally stored.