Agentic Compute MCP server

🏆 Track 01 Hackathon Criteria Addressed

This project strictly adheres to the Razorpay Track 01 requirements for AI-agent storefronts:

🔒 Gated (Transactable End-to-End):

AI Agents are charged per API call. If the wallet is empty, the FastAPI middleware intercepts the request and throws a graceful 402 Payment Required. If funded, it verifies the crypto token and settles USDC via the Base network.

⏱️ Bounded (Zero Resource Exhaustion):

All sandboxes run inside Azure Container Apps Dynamic Sessions. Code execution is strictly bounded to a 15-second maximum timeout limit enforced by Pydantic models. Malicious memory-clogging scripts are instantly killed.

📊 Explainable (Audit Trail):

Every successful and failed transaction generates an immutable AUDITRECORD locally, detailing the exact priceusdc, duration_ms, and execution exit code.

🏗️ Architecture & File Structure

graph TD
    A[Claude Desktop App<br>Agent Client] -->|Tool Invocation| B{FastAPI x402 Middleware}
    
    subgraph Gate ["The Payment Gate"]
        B -->|Funded Wallet| D[Settle USDC via Base Network]
        B -->|Empty Wallet| C[HTTP 402 Payment Required<br>Graceful Failure]
        D -.->|Log| Z[(Local Audit Trail)]
        C -.->|Log| Z
    end
    
    %% Invisible structural link to force vertical placement
    Z ~~~ E
    
    D --> E[Azure Container Apps<br>Dynamic Session Pool]
    C -.->|Agent Retries| A
    
    subgraph Cloud ["Bounded Cloud Execution"]
        E --> F[15s Python Execution Timeout]
        F --> G[Generate Text / Base64 Image]
    end
    
    G -->|Stream Payload| H[mcp_client.py]
    
    subgraph Local ["Local System (Zero-Context Rendering)"]
        H -->|JSON Response| A
        H -->|Image Payload| I[(Save to Local PNG)]
    end
    
    classDef gate fill:#ffebee,stroke:#c62828,stroke-width:2px;
    classDef success fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px;
    class C gate;
    class D success;

agentic-compute-mcp/
├── README.md                      # Documentation & Setup
├── .env.example                   # Environment variable template
├── main.py                        # The Backend: FastAPI, x402 Gate, and Azure routing
├── mcp_client.py                  # The Bridge: Claude Desktop tool definitions
├── sandbox.py                     # Core Azure Dynamic Sessions execution logic
├── pyproject.toml                 # Python package and dependency configurations
├── requirements.txt               # Dependencies (FastAPI, azure-identity, uvicorn, etc.)
├── server.json                    # MCP server configuration details
├── glama.json                     # Glama registry metadata for the AI agent storefront
└── llms.txt                       # Context file for LLM integration

🖼️ Feature Highlight: Zero-Context Image Rendering

Handling base64 image strings in LLM context windows is notoriously unreliable and eats up thousands of tokens. agentic-compute-mcp solves this locally.

When Claude calls generateplot, the Azure sandbox generates the chart and streams the payload back to the MCP client. The client automatically intercepts the payload, decodes it, and saves it directly to your local machine as optimizedload_trend.png—completely bypassing the LLM context window to prevent token exhaustion.

🌙 Developer Note: The 2 AM Story & Graceful Failures

The Graceful Failure: To meet the Track 01 requirement for agent-to-agent commerce, the backend is designed to fail gracefully. If an agent attempts to execute code without funding, the FastAPI middleware intercepts the payload and throws a clean 402 Payment Required error. This prevents compute theft and allows the agent to automatically reroute to the /verify and /settle endpoints.

Shortened. The full README is on GitHub.