Claude Desktop

• Claude Chat, which provides general conversation with no file access, similar to claude.ai. At the time this article was initially published, Claude Chat was available on all Claude pricing plans.
• Claude Cowork, an autonomous background agent that works on tasks in a cloud virtual machine with its own environment. It can run independently while you do other work. At the time this article was initially published, Claude Cowork was available only Claude Pro, Max, Team, and Enterprise pricing plans.
• Claude Code, an interactive coding assistant with direct access to your local files. You review and approve each change in real time. At the time this article was initially published, Claude Code was available only on Claude Pro, Max, Team, and Enterprise pricing plans.

​

Requirements

1. Sign up for a Claude paid plan or upgrade from a free to a paid Claude plan, if you have not already done so.
2. Install Claude Desktop on your local machine, if you have not already done so.
3. Create an Unstructured workflow that connects one or more supported sources containing your organization’s source documents, semi-structured data, or both, to a supported destination that will contain Unstructured’s AI-ready outputs based on those source documents, semi-structured data, or both. This article has been tested with the following Unstructured destination types:
   • DataStax Astra DB
   This list will be updated over time as additional Unstructured destination types are tested. Any supported Unstructured destination type that has a corresponding available Model Context Protocol (MCP) server should work. Most of these MCP servers are available through third parties and not through Unstructured.
   MCP servers that are offered by third parties are neither owned nor supported by Unstructured. For questions, issues, or feature requests for a specific MCP server, contact that MCP server’s provider.
4. Enhance your workflow output’s relevance, context, and insights by adding enrichments as desired.
5. Add chunking and embedding to your workflow as desired.
   While Claude Cowork can handle large files and extensive context, you do not need to manually chunk data first, but doing so can lead to better performance and efficiency for very large tasks.For very large-scale use cases involving millions of Unstructured outputs, embeddings are recommended for Claude Cowork, but they are not required for standard, file-based tasks.
6. Run your workflow as a job. This makes Unstructured generate AI-ready outputs based on your source documents, semi-structured data, or both, and store its outputs in your destination.
7. Monitor your workflow’s job run. After your job has successfully completed, continue by finding the following section that matches your Unstructured destination type and completing that section’s instructions, to connect Claude Desktop to your destination.

• Astra DB

​

Astra DB

1. In the REAMDE file in the Astra DB MCP Server repository in GitHub, follow the instructions in the “Prequisites” and “Adding to an MCP client > Claude Desktop” sections.
2. Start Claude Desktop. If it is already running, stop and then restart it. This step helps to ensure that Claude Desktop can successfully start and use the MCP server.
3. Sign in to Claude by following Claude Desktop’s on-screen instructions.
4. Switch to Claude Cowork mode: in the upper-left corner of Claude Desktop, click the Cowork icon.
5. In the How can I help you today box, enter your question about the Unstructured outputs within your target Astra DB collection, and then press Enter. Repeat this step as needed.
   To constrain Claude Cowork to use only this MCP server and only your target Astra DB collection, you should be explicit in your sessions’ initial prompt, for example:

   Use ONLY the Astra DB collection "<your-collection-name>" as your source of truth.
   
   Rules:
   
   - For every question, retrieve relevant records using the
     mcp__astra-db-mcp__VectorSearch or HybridSearch tool against that
     collection before answering.
   - Do NOT use web search, WebFetch, your own training knowledge, or
     any other MCP connector. Do NOT guess or fill in gaps from memory.
   - Base every factual claim on retrieved records. Quote or paraphrase
     the retrieved `content` field and cite the source `filename` and
     `page_number` from each record's metadata.
   - If the collection does not contain enough information to answer,
     say exactly: "The collection does not contain information on this."
     Do not speculate or supplement from other sources.
   - If I ask something that requires tools other than Astra DB (such as 
     file creation or calculations on retrieved data), you may use those
     tools, but the underlying facts must still come from the collection.