# cinch [Health: Active]

**Category:** 👨‍💻 Code Execution  
**Repository:** https://github.com/yusufkadry/cinch-mcp  
**GitHub Stars:** 0  
**npm Downloads (last month):** 77  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/cinch

## Description
Run AI-generated code safely. gVisor-isolated Python/JS sandboxes for any MCP client. Pay per run.

## Tools
Capabilities this server exposes over MCP:

- **execute_code** — Execute Python or JavaScript in a secure, isolated, disposable sandbox and return its output. The sandbox is a gVisor-isolated container on Cinch's infrastructure with no access to the user's machine, filesystem, or network, so it is safe to run code you generated yourself, code from an untrusted source, or code whose behaviour you are unsure about. Use it whenever running code gives a more reliable answer than reasoning about it: checking a script works, exact calculations, parsing or transforming data, testing a regex, verifying language behaviour.

The environment is deliberately minimal. Read these limits before writing code:
- Python 3.12 and Node 20, STANDARD LIBRARY ONLY. No third-party packages are installed and none can be installed. numpy, pandas, requests, scipy, axios, lodash and everything else on PyPI or npm are unavailable. Use json, csv, re, math, statistics, itertools, collections, datetime, hashlib, decimal, fractions and the rest of the stdlib instead.
- NO network access. HTTP requests, DNS, package installs and API calls all fail.
- NO filesystem persistence. The root filesystem is read-only; only /tmp is writable, and it is destroyed when the run ends.
- 10 second execution limit, 256 MB memory, 0.5 CPU.
- Every call gets a brand new sandbox. Nothing persists between calls — no variables, no files, no imports. Each snippet must be complete and self-contained.
- Only stdout and stderr are returned. Print anything you want to see; a bare expression on the last line returns nothing.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "cinch": {
    "command": "npx",
    "args": ["-y","@cinch-codes/mcp"],
    "env": {
      "CINCH_API_KEY": "",
      "CINCH_TIMEOUT_MS": "",
      "CINCH_BASE_URL": ""
    }
  }
}
```

**Requires environment variables:** `CINCH_API_KEY`, `CINCH_TIMEOUT_MS`, `CINCH_BASE_URL` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What cinch MCP server does

The cinch MCP server exposes Cinch’s remote code-execution service to MCP clients. An assistant sends a complete Python or JavaScript program through the `execute_code` tool, and receives stdout, stderr, the exit code, and execution duration. Programs must print results to stdout if the assistant needs to read them.

Execution happens on Cinch infrastructure rather than on the host running the MCP client. The container uses gVisor isolation, runs as a non-root user, drops capabilities, and has no access to the client machine, local files, or local network. Cinch charges per execution and does not require a subscription minimum.

## How it works

The cinch MCP server uses a stdio process launched through `npx`. The client starts `@cinch-codes/mcp`, and the server sends execution requests to the Cinch API using the configured API key. `npx` downloads the package on its first run, so a separate package installation is not required.

Each call starts with a clean environment. State does not carry from one request to the next, so code should be self-contained. Python 3.12 and Node 20 are available, with standard-library modules only. The sandbox has no DNS or HTTP access, a read-only root filesystem, a writable 64 MB `/tmp`, 256 MB of memory, 0.5 CPU cores, and a 10-second execution limit.

## Setup and configuration

Create a Cinch API key, then add the server to an MCP client. Claude Desktop and Cursor use an MCP server entry with `command` set to `npx`, arguments `-y` and `@cinch-codes/mcp`, and `CINCH_API_KEY` in the environment. Claude Code can add the same process with `claude mcp add`. On Windows, wrap `npx` with `cmd /c`.

The required variable is `CINCH_API_KEY`. `CINCH_TIMEOUT_MS` sets the client-side timeout and defaults to 20,000 milliseconds; Cinch still caps execution at 10 seconds. `CINCH_BASE_URL` optionally changes the API endpoint, which defaults to `https://api.cinch.codes`. Node.js 18 or newer is required.

## Tools and capabilities

The cinch MCP server provides one tool:

- `execute_code`: runs a complete program in either Python or JavaScript. The `language` value defaults to Python.

Output from each stream is limited to 20,000 characters and is truncated beyond that point. Timeouts and non-zero exits are reported as tool errors, allowing the assistant to recognize failed runs. Responses for exhausted credits identify the condition and tell the assistant not to retry.

## Limitations and notes

This environment is intentionally minimal. Packages cannot be installed during execution, and code that depends on third-party Python or npm modules will not work. Network-dependent programs also fail. Temporary files disappear when the run ends, and the root filesystem cannot be modified.

Windows clients may report connection failures if `npx` is not launched through `cmd /c`. If `npx` reports a missing npm directory, create `%APPDATA%\npm` and retry. When diagnosing startup problems, run the process directly with `CINCH_API_KEY` set; a ready message followed by silence indicates that the MCP server is waiting for a client.

_Full upstream README: https://allmcps.com/mcp/cinch/readme_