# Tamperbox MCP

A code-mode MCP server with two tools over a documented catalog of Tamperbox operations.

Canonical: https://docs.tamperbox.dev/mcp/
Updated: 2026-10-19
Applies to: MCP 0.2.0
Prerequisites: An MCP-compatible client
Outputs: A clear choice between bundled search and API-backed execution
Limitations: The MCP catalog covers API operations, not every interactive browser mode.

Tamperbox MCP exposes two tools: `search` and `execute`. Both accept an async JavaScript function body. The smaller interface lets agents inspect exact schemas before they send inputs to the API.

## Tool model

### `search`

Runs against the bundled OpenAPI document. It exposes `spec` and a resolved `operations` array. It does not call the Tamperbox API, and `apiCalls` stays at `0`.

### `execute`

Runs code in an isolated QuickJS WebAssembly environment. The code can call documented endpoints through `tamperbox.request({ method, path, query?, body? })`.

```js
const result = await tamperbox.request({
  method: 'POST',
  path: '/v1/tools/dns_lookup',
  body: { name: 'example.com', type: 'A', resolver: 'google' },
});
return result;
```

## Transports

The package supports stdio for local clients and Streamable HTTP for hosted integrations. The public production endpoint is `https://mcp.tamperbox.dev/mcp`.

## What it does not do

The server cannot call arbitrary URLs, add arbitrary headers, publish DNS changes, send email, or establish JWT trust through decoding. Unknown routes and arguments are rejected against the bundled contract.