Azure Cosmos DB now has an official MCP toolkit, and it hit general availability this week. If you build AI agents that need to talk to databases, this is worth a look. The Model Context Protocol (MCP) is an open standard from Anthropic that lets AI models interact with external tools and data sources through a common interface. Think of it as USB-C for LLM tool calling: one protocol, any compatible server. Microsoft has been shipping MCP servers for its services at a steady clip, and Cosmos DB is the latest to get first-class treatment.
What the toolkit actually does
The Cosmos DB MCP toolkit is an open-source server that sits between your AI agent and your Cosmos DB instance. Your agent sends a natural language request — “show me orders from last week where the status is pending” — and the MCP server translates that into a Cosmos DB query, runs it, and returns structured results. No custom API layer. No hand-rolled SQL generation.
At MS Build 2026, Microsoft also announced semantic reranking support for the toolkit. This means the server can use the vector search capabilities in Cosmos DB to reorder results by relevance to the query intent, not just keyword match. For RAG applications, that matters a lot.
The toolkit is provider-agnostic within the MCP ecosystem. If your agent already speaks MCP — whether it is Claude Desktop, a custom LangChain agent, or something built on the MCP SDK — adding Cosmos DB is a config change, not a code rewrite.
Setting it up
Installation is straightforward if you are already in the Azure and Node.js ecosystem:
- Clone the GitHub repo
- Configure your Cosmos DB connection string and database name
- Point your MCP-compatible agent at the server endpoint
The repo includes a VS Code extension integration and samples for common agent frameworks. You can run it locally during development or deploy it as a container alongside your agent infrastructure.
Where it fits (and where it does not)
This is not a replacement for your application’s data access layer. The MCP toolkit is designed for agent-driven, ad-hoc queries — the kind where the user asks a question and the agent needs to pull live data to answer. It is not optimized for high-throughput OLTP workloads or reporting pipelines.
Security is the obvious concern. Giving an AI agent query access to a production database is not something you do casually. The toolkit relies on the MCP authentication model and Cosmos DB’s built-in RBAC. You should run it with a restricted service principal, not an admin key. Microsoft’s docs recommend using Entra ID with least-privilege roles, and that is not optional advice if you are pointing this at real data.
Also worth noting: the Cosmos DB MCP toolkit is one of several. Azure DocumentDB got its own MCP toolkit in preview at the same time. They share the same protocol but target different database engines. If you are on DocumentDB, use that one.
If you are already building agents that query structured data, this removes a chunk of integration work. If you are not, it is a signal of where Microsoft thinks agent architecture is going: standardized protocols, not bespoke REST wrappers.