# agentic-compute [Health: Active]

**Category:** 👨‍💻 Code Execution  
**Repository:** https://github.com/codelad1304/agentic-compute-mcp  
**GitHub Stars:** 0  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agentic-compute

## Description
Secure Python sandbox for data optimization and Matplotlib rendering via x402 microtransactions.

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What agentic-compute does

The agentic-compute MCP server exposes four paid tools for data processing and Python execution. `execute_code` accepts a Python script and returns captured output, with an 8 KB response limit. `sanitize_csv` converts an unformatted CSV string into a JSON array, including null handling, header normalization, and removal of empty rows.

`optimize_ga` accepts numerical data and returns model parameters with a MAPE score for forecasting, curve fitting, or related optimization tasks. `generate_plot` accepts coordinates and chart configuration, then saves the resulting chart locally through the MCP client. The documented tool prices are 0.10 USDC for code execution, 0.25 USDC for CSV sanitization, 0.50 USDC for genetic-algorithm optimization, and 0.30 USDC for plot generation.

## How it works

The agentic-compute MCP server uses `mcp_client.py` as the bridge between an MCP host and the FastAPI backend. The backend applies x402 payment middleware before routing execution requests to Azure Container Apps Dynamic Sessions. An unfunded request receives HTTP 402 Payment Required; funded requests verify the crypto token and settle USDC through the Base network.

Each execution runs in a fresh Azure Dynamic Session with a documented maximum boundary of 15 seconds. The project describes these sessions as isolated microVMs and notes that fresh provisioning can add cold-start latency. Successful and failed transactions are recorded locally with the USDC price, duration, and execution exit code.

For plotting, the Azure sandbox returns the generated payload to the client. The client decodes the image and writes it to local storage, avoiding the need to place base64 image data in the model context. The README identifies `optimized_load_trend.png` as the generated filename for this workflow.

## Setup and configuration

Install the Python dependencies from `requirements.txt`, create a `.env` file from `.env.example`, and configure the Azure pool endpoint and wallet address. Start the backend with Uvicorn on the documented local address and port. The MCP client is configured separately in Claude Desktop using a `python` command whose argument points to `mcp_client.py`.

The client configuration also requires an EVM private key and a funded wallet for x402 microtransactions. Keep that key out of source control. The backend and client are separate components, so both must be configured for tool calls to reach the Azure execution pool.

## Tools and capabilities

- Execute arbitrary Python in the remote sandbox; do not use this tool for local filesystem operations.
- Clean raw CSV strings before analysis by normalizing headers, converting NaN values to null, and dropping empty rows.
- Optimize numerical data and return parameters with a MAPE result.
- Generate charts without requiring local GUI dependencies, saving the decoded PNG through the client.

## Limitations and notes

The agentic-compute MCP server is not a free local execution service: every tool invocation has a listed USDC price and requires payment configuration. Code output is limited to 8 KB, and execution is bounded at 15 seconds. Empty wallets produce a payment error rather than allowing unpaid execution.

Azure Dynamic Sessions can introduce several seconds of startup time because a fresh environment is provisioned for each request. Plot output is intended to be consumed from the locally saved file rather than read back as base64 data by the model. The repository is released under the MIT License.

_Full upstream README: https://allmcps.com/mcp/agentic-compute/readme_