Skip to main content

Unstract Platform MCP server

Unstract has support for MCP server which can be used to integrate the functionality of connecting to an API deployment with a MCP host.

Features​

Connects to a deployed Unstract API to process files and extract structured data. Handles asynchronous processing and polling for completion.

Tools​

unstract_tool: Submits a file to the Unstract API, polls for completion, and returns structured extraction results. Handles errors, timeouts, and environment-based configuration. The tool also supports optional paramters to include metadata and metrics.

Configuration​

Getting an API Key​

To use the tool, one needs to set the UNSTRACT_API_KEY and API_BASE_URL environment variables. For obtaining these, please check the details under Getting Started

Usage with Claude Desktop​

One can integrate the Unstract tool with Claude Desktop by adding the following configuration to the claude_desktop_config.json replacing the values of UNSTRACT_API_KEY and API_BASE_URL with the right values. Also, since the tool will need a file to be processed, the file needs to be mounted in the container. The sample below uses /tmp as the mount point. This can be changed to any other location.

Docker​

{
  "mcpServers": {
        "unstract_tool":{
            "command": "/usr/local/bin/docker",
            "args": [
                "run",
                "-i",
                "--rm",
                "-v",
                "/tmp:/tmp",
                "-e",
                "UNSTRACT_API_KEY",
                "-e",
                "API_BASE_URL",
                "unstract/mcp-server",
                "unstract"
            ],
            "env": {
                    "UNSTRACT_API_KEY": "",
                    "API_BASE_URL": "https://us-central.unstract.com/deployment/api/...../"
            }
        }
    }
}

Once the configuration is added, user can use extract_text as a tool to extract text from the supported documents.

Sample prompt​

Process the document /tmp/sample-bank_statement.pdf

This should invoke the tool and get the response from the API deployed at the mentioned location. Optionally, one can request for includeing metadata and metrics also.