Tensorlake MCP Server

An MCP server that provides a cloud sandbox environment for document processing. Upload documents, parse them with AI, and work with the results using familiar harness engineering tools (bash, file read/edit, grep, glob).

Prerequisites

• A Tensorlake API key (sign up at Tensorlake)
• An MCP-compatible host application (e.g., Claude Desktop, Cursor, or other MCP hosts)

Installation

Install using go install

go install github.com/sixt/tensorlake-mcp@latest

Locate the binary in your GOPATH/bin directory:

$ which tensorlake-mcp
/Users/<username>/go/bin/tensorlake-mcp

Building from Source

git clone https://github.com/sixt/tensorlake-mcp.git
cd tensorlake-mcp
go build -o tensorlake-mcp .

Configuration

Environment Variables

MCP Host Setup

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "tensorlake-mcp": {
      "command": "/absolute/path/to/tensorlake-mcp",
      "env": {
        "TENSORLAKE_API_KEY": "your-api-key-here",
        "TENSORLAKE_SANDBOX_API_BASE_URL": "https://api.tensorlake.ai/sandboxes",
        "TENSORLAKE_SANDBOX_PROXY_BASE_URL": "https://sandbox.tensorlake.ai"
      }
    }
  }
}

Claude Code

claude mcp add tensorlake-mcp /absolute/path/to/tensorlake-mcp \
  -e TENSORLAKE_API_KEY=your-api-key-here \
  -e TENSORLAKE_SANDBOX_API_BASE_URL=https://api.tensorlake.ai/sandboxes \
  -e TENSORLAKE_SANDBOX_PROXY_BASE_URL=https://sandbox.tensorlake.ai

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "tensorlake-mcp": {
      "command": "/absolute/path/to/tensorlake-mcp",
      "env": {
        "TENSORLAKE_API_KEY": "your-api-key-here",
        "TENSORLAKE_SANDBOX_API_BASE_URL": "https://api.tensorlake.ai/sandboxes",
        "TENSORLAKE_SANDBOX_PROXY_BASE_URL": "https://sandbox.tensorlake.ai"
      }
    }
  }
}

See also: Other MCP Clients

Tools

All tools operate on a cloud sandbox filesystem. The sandbox is created lazily on first use and all files live under /data. The sandbox session is persisted across server restarts — if the sandbox is still running, it will be reused automatically.

The server instructs AI clients to prefer dedicated tools over bash for file operations, following harness engineering best practices.

bash

Execute shell commands in the sandbox. Returns exit code, stdout, and stderr. Sends MCP progress notifications during long-running commands.

bash_status

Check status or retrieve results of a background bash command.

file_read

Read a file from the sandbox with line numbers (cat -n format). Detects binary files, returns base64 image content for image files, and supports PDF text extraction with page ranges.

file_write

Write a file to the sandbox. Creates or overwrites entirely.

file_edit

Edit a file using exact string replacement.

grep

Search file contents by regex pattern. Supports multiple output modes, context lines, and case-insensitive search.

glob

Find files by name pattern. Supports recursive ** patterns. Results sorted by modification time.

upload

Upload a file into the sandbox from a URL, local path, or data URI.

parse

Parse a document using Tensorlake AI. Writes the result as markdown.

Example Interaction

User: Parse this invoice at https://example.com/invoice.pdf and find the total amount.

AI: [Uses upload to fetch the PDF into the sandbox]
    [Uses parse to extract content as markdown]
    [Uses file_read to inspect the parsed result]
    [Uses grep to search for the total amount]
    The total amount on the invoice is $1,234.56.

Development

export TENSORLAKE_API_KEY="your-api-key"
go build && ./tensorlake-mcp

Testing

Integration tests require a valid API key:

TENSORLAKE_API_KEY="your-api-key" go test -v ./...

MCP Inspector

npx @modelcontextprotocol/inspector /absolute/path/to/tensorlake-mcp

License

Copyright 2025 SIXT SE. Licensed under the Apache License, Version 2.0. See LICENSE for details.