The Agentic Data Stack

One of the first things we tried when we got access to an LLM was to ask it to generate SQL. In the early days the results were unreliable — the models struggled to interpret intent and often produced SQL that was syntactically or semantically incorrect.

Over the last couple of years we have watched the models improve dramatically. In a bake off between Claude and Gemini, we could see that text to SQL was almost ready for prime time. In parallel, agents have been on a similar journey — LLMs which can plan, iterate, call tools and correct themselves have moved from a novelty to something genuinely useful.

These two worlds are now converging. When you combine reliable text to SQL with an agent that can decompose a question, run intermediate queries and inspect the results, you end up with something that starts to look like a real replacement for traditional business intelligence. We have written more about what this means in practice in What Do we Mean By Agentic Analytics.

The Agentic Data Stack

ClickHouse have packaged the ingredients you need to build this style of solution into what they refer to as the agentic data stack. It has three main components:

• ClickHouse as the underlying real time database. As we explain in Why ClickHouse Is The Best Data Platform For Agentic AI, the combination of low latency, high concurrency, freshness and price/performance makes it uniquely well suited to the workloads agents create.
• The ClickHouse MCP server, which exposes the database to LLMs through the emerging Model Context Protocol standard. This gives the agent a well defined interface for discovering schema, running queries and returning results without a lot of bespoke plumbing.
• LibreChat, a recent ClickHouse acquisition which provides the frontend. LibreChat lets you connect to any model provider (Anthropic, OpenAI, self hosted, OpenRouter, Bedrock), configure MCP servers, run code artefacts for visualisation, and define named agents with prompts and tool access.

Together these give you an end to end path from a business question typed into a chat interface, through an agent that reasons and plans, into ClickHouse for the underlying data, and back to a visualisation or narrative answer.

Building Agent Teams

One of the interesting patterns we are seeing with clients is the move from a single generic assistant to a team of specialised agents. You might deploy an HR agent, a finance agent, a data analyst agent and a compliance agent — each configured with a focused prompt, its own tool access and its own permissions on the underlying data.

LibreChat is well suited to this style of deployment. You define the agent once, package it up and expose it to the business as a tool that colleagues can draw on in their day to day work. Over time this becomes a genuinely differentiated capability that is hard to replicate with off the shelf products.

Why This Combination Works

There are a few reasons this stack is compelling in practice.

The first is that it is quick to stand up. ClickHouse runs as a single binary in open source or as a managed service in the cloud. LibreChat runs alongside it with the MCP server configured in a few lines. With very little effort you have a working database, MCP layer and frontend that you can start pointing at real data.

The second is that it is open and unopinionated about models. You are not tied to a single LLM provider — you can use frontier models where you need the intelligence, or open source and open weight models where cost, sovereignty and privacy matter more. We discuss this trade off in more detail in Using Open Source and Open Weight LLMs With ClickHouse and Self Hosting An Open Source Model With ClickHouse.

The third is that ClickHouse itself is the right foundation. If you deploy agents at scale, a single question can turn into tens of concurrent queries. If the business achieves its long standing aim of a data driven culture and expose agents to end customers too, this can quickly become thousands of concurrent queries. The underlying database has to stand up to that workload at an acceptable price point, and that is exactly what ClickHouse is designed for.

Demo: Capital Markets Use Case

Here is a demo we put together of the agentic data stack applied to a capital markets use case:

Wrapping Up

The agentic data stack is not a single product so much as a curated set of components that fit together well. ClickHouse provides the data platform, the MCP server provides the interface, and LibreChat provides the user experience and agent orchestration layer.

For businesses looking to move beyond static dashboards and towards genuine conversational analytics and agent teams, this is one of the most credible starting points we have seen. It is quick to deploy, open at every layer, and built on a database that will scale with the workload as adoption grows.