The Agentic Data Stack

See below for a demo of the ClickHouse agentic data stack.

The Evolution of Text-to-SQL

One of the first things I tried when I got access to an LLM was to use it to generate SQL. In the early days, the models were unreliable, both in building semantically correct SQL and understanding your intent.

However, I've watched the models improve over the last couple of years, and last year I ran a series of tests where I gave it SQL challenges where I did a bake off between Claude and Gemini. I could see that the text to SQL was almost ready for prime time.

In parallel, I've been building more agents. This is where you give an LLM access to tool calls such as APIs, databases, MCP servers, and documentation. Likewise, these have also been on a journey of improvement as evidenced in how people are using coding agents.

These worlds are now combining, whereby we have a good base of text to SQL and agents which can think through and iterate on problems and make use of tools to provide very powerful and insightful data analysis and generative business intelligence.

The Agentic Data Stack

ClickHouse has packaged all of the tools into the agentic data stack, which is a combination of ClickHouse to hold the underlying data, the MCP server to make it easier for an LLM to work with ClickHouse and LibreChat, which was one of their recent acquisitions.

LibreChat gives us a frontend for interacting with LLMs. We can choose from all of the different providers, such as Anthropic or OpenAI. We can use local LLMs, or we can use routers such as OpenRouter or AWS Bedrock. It has support for MCP servers such as the ClickHouse MCP. It has support for artifacts, which is where we can run code that gives us visualizations of many apps. There is also an Agent Experience where we can define agents, give them a prompt, give them access to tools, and give them permission to use things like MCPs and things like that. We can then package up the agent and expose it to the business as a tool to help people in their day-to-day jobs.

Building Agent Teams

The direction this seems to be going is that you might deploy an HR agent, a finance agent, a data analyst agent, maybe a compliance agent, and things like that. They are then there to draw upon, all configured with the correct tools.

ClickHouse has given us the ability to stand up with Stack in a very simple manner. ClickHouse is easy to run either in the cloud or as a single binary open source. We can then run LibreChat which also combines the MCP server with a little configuration. With a few commands, we essentially have this end-to-end stack where we have a database, we have an MCP, and we have a front end.

Demo: Capital Markets Use Case

Here is a demo I put together of the Agentic Data Stack for a capital markets use case: