The cornerstone of autonomous AI Agentic systems and GenAI applications is an “Augmented LLM”, which is defined as an LLM enhanced with augmentations from various data sources and knowledge bases. Since its introduction, there has been a lot of buzz around Model Control Protocol (MCP), an evolving protocol standard that specifies how applications can provide context to AI Models/LLMs. Using the context, AI agents can autonomously decide which tools to use in order to accomplish a task.

Today, we are happy to introduce the Couchbase MCP Server for Couchbase that can be leveraged with AI agentic workflows and applications by enabling LLMs to perform actions against your Couchbase cluster through a well-defined set of tools. The Couchbase server  data source may be hosted on Capella or self managed.

How to get started

The MCP Server is an open source, Couchbase Community Supported offering that is currently available as a self-managed offering with a fully managed version coming later. The server is available on several of the popular destinations for MCP servers and more to come:

The current version of MCP server must be configured with a specific bucket in the Couchbase Server cluster, which may be self managed or hosted on Capella.  All data operations are in the context of that bucket. The version of MCP Server at the time of this blog supports the following tools:

    • Get_scopes_and_collections_for_bucket
      • Fetches a list of all the scopes and collections in a bucket.
    • Get_schema_for_collection
      • Fetches the data model associated with documents in a specific collection
    • Run_sql_plus_plus_query 
      • Runs SQL++ query against specific bucket. Currently, read-only query operations are permitted.
    • read_document
      • Retrieves document matching specific criteria
    • add_document
      • Inserts a document
    • delete_document
      • Deletes document
    • update_document
      • Replaced current version of document with new document

With new tools and updates coming in on a regular basis, ensure you are working with the latest version. Support for resources and prompts are planned for the future.

Try it!

You will likely be building your own MCP-aware AI agents and applications that will be leveraging Couchbase MCP server but the standard protocol allows you to quickly get started with evaluating  the server using one of the many MCP clients, including Claude desktop, cursor and Windsurf.

See it in action

This video recording is a simple demonstration of Couchbase MCP Server that is configured with my Capella free tier cluster as backing data source.

    • The cluster is provisioned with some rudimentary store product catalog data set.
    • Using Claude desktop as the client, we ask questions in natural language against the Claude3.7 Sonnet model, which in turn uses MCP to talk to the Couchbase MCP server for retrieving additional context.
    •  In the video, you can see that the LLM determines the tools to use and the order of execution in order to accomplish the task.
    • Once the task is completed, the results are verified by running SQL++ query directly on the cluster.

Next Steps

We’d love to hear from you. If you have questions, or feedback on a tool you’d like, you can reach out to us via Couchbase Forums, Discord or GitHub.

Acknowledgements

Special thanks to Nithish for his contributions to the project.

Author

Posted by Priya Rajagopal, Senior Director, Product Management

Priya Rajagopal is a Senior Director of Product Management at Couchbase responsible for developer platforms for the cloud and the edge. She has been professionally developing software for over 20 years in several technical and product leadership positions, with 10+ years focused on mobile technologies. As a TISPAN IPTV standards delegate, she was a key contributor to the IPTV standards specifications. She has 22 patents in the areas of networking and platform security.

All Posts

Leave a reply