cinch-mcp

# @cinch-codes/mcp

MCP server for [Cinch](https://cinch.codes). Gives your AI assistant a real sandbox to run code in.

Without it, an assistant writes code and you run it yourself. With it, the assistant runs the code and reads the actual output — inside a gVisor-isolated container on Cinch's infrastructure, with no access to your machine, filesystem, or local network.

Pay per execution. No subscription floor.

## Setup

Get an API key at [cinch.codes](https://cinch.codes), then add the server to your MCP client.

**Claude Desktop** — `claude_desktop_config.json`:

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

**Claude Desktop on Windows** — same file, but Windows can't spawn `npx` directly, so wrap it with `cmd /c`:

```json
{
  "mcpServers": {
    "cinch": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "@cinch-codes/mcp"],
      "env": {
        "CINCH_API_KEY": "cinch_live_..."
      }
    }
  }
}
```

**Claude Code** (macOS/Linux):

```bash
claude mcp add --scope user cinch -e CINCH_API_KEY=cinch_live_... -- npx -y @cinch-codes/mcp
```

**Claude Code** (Windows):

```cmd
claude mcp add --scope user cinch -e CINCH_API_KEY=cinch_live_... -- cmd /c npx -y @cinch-codes/mcp
```

**Cursor** — `.cursor/mcp.json`, same shape as the Claude Desktop config above.

Restart the client. That's it — no install step, `npx` fetches it on first run.

## What it exposes

### `execute_code`

Runs a self-contained Python or JavaScript program and returns stdout, stderr, exit code, and duration.

| Parameter | Type | Default | Description |
|---|---|---|---|
| `code` | string | — | The complete program to run. Must print to stdout to return anything. |
| `language` | `python` \| `javascript` | `python` | Runtime to execute in. |

Each call gets a clean sandbox. State does not persist between calls, so every snippet needs to stand on its own.

## The sandbox environment

Deliberately minimal. Worth knowing before you wonder why an import failed:

| | |
|---|---|
| Runtimes | Python 3.12, Node 20 |
| Packages | **Standard library only.** No pip or npm packages are installed, and none can be installed at runtime. |
| Network | **None.** HTTP, DNS, and package installs all fail. |
| Filesystem | Root is read-only. `/tmp` is writable (64 MB) and destroyed when the run ends. |
| Memory | 256 MB |
| CPU | 0.5 cores |
| Time limit | 10 seconds |
| Isolation | gVisor (`runsc`), all capabilities dropped, no-new-privileges, non-root user, 64-process cap |

The tool description tells the model all of this up front, so it writes stdlib-only code instead of reaching for `numpy` and failing on the first call.

## Configuration

| Variable | Required | Default | Description |
|---|---|---|---|
| `CINCH_API_KEY` | yes | — | Your Cinch API key. |
| `CINCH_TIMEOUT_MS` | no | `20000` | Client-side timeout in ms. The API caps execution at 10s regardless. |
| `CINCH_BASE_URL` | no | `https://api.cinch.codes` | Override the API endpoint. |

## Troubleshooting

**Windows: "Failed to connect" in `claude mcp list`** — you're missing the `cmd /c` wrapper. Windows resolves `npx` to a batch script that can't be spawned directly; re-add the server using the Windows command above.

**Windows: `npm error ENOENT ... AppData\Roaming\npm`** — some Node installs never create npm's global folder, and `npx` refuses to run without it. Create it once and retry:

```cmd
mkdir %APPDATA%\npm
```

**Server exits immediately with "CINCH_API_KEY is not set"** — the env var didn't reach the server. In Claude Code, put `-e CINCH_API_KEY=...` *before* the server name in `claude mcp add`. In config files, check the `env` block is inside the `cinch` entry.

**Debugging any connection failure** — run the server directly to see the real error instead of a generic status:

```bash
CINCH_API_KEY=cinch_live_... npx -y @cinch-codes/mcp   # macOS/Linux
```
```cmd
set CINCH_API_KEY=cinch_live_... && cmd /c npx -y @cinch-codes/mcp   # Windows
```

Correct behavior is `cinch-mcp ... ready` followed by silence — an MCP server waits for a client. Anything else printed is the actual failure.

## Why sandboxed execution

Code written by a model is untrusted code — nothing reviewed it before it ran. Executing it directly on your machine means handing it your filesystem, your network, and your credentials. Cinch runs it somewhere else entirely, in a disposable container with kernel-level isolation, and sends back only the output.

## Notes

- Requires Node 18 or newer.
- Output is capped at 20,000 characters per stream to protect your context window; anything beyond that is truncated with a marker.
- Runs that time out or exit non-zero are returned as tool errors, so the assistant knows the code failed and can correct it.
- Out-of-credit responses are surfaced clearly and instruct the assistant not to retry, so a drained balance does not turn into a retry loop.

## Links

- [cinch.codes](https://cinch.codes)
- JS/TS SDK: [`@cinch-codes/pangolin`](https://www.npmjs.com/package/@cinch-codes/pangolin)
- Python SDK: [`pangolin-sdk`](https://pypi.org/project/pangolin-sdk/)

MIT