Agentic Data Stack

The open-source stack for ClickHouse's suite of agentic analytic tools — your chat, your models, your data.
Powered by ClickHouse, LibreChat, and Langfuse.

Learn more at clickhouse.ai and the Agentic Data Stack blog.

One Click Deploy

Want to skip the setup and jump straight into adding agentic analytics into your workflow? Choose a Railway template below to spin up a preconfigured instance of the Agentic Data Stack in the cloud.

Langfuse OSS	Langfuse Cloud

More details available in the Deploy on Railway section.

Overview

This project runs a fully self-hosted agentic analytics environment with Docker Compose. It connects a chat UI (LibreChat) to your data (ClickHouse) via MCP, with full LLM observability (Langfuse) — all in a single docker compose up command.

What's included

Component	Purpose	Port
LibreChat	Modern Chat UI with multi-model / provider support (OpenAI, Anthropic, Google)	3080
Admin Panel	Browser-based configuration UI for LibreChat	3081
ClickHouse MCP	MCP server that gives agents access to ClickHouse	8000
Langfuse	LLM observability — traces, evals, prompt management	3000
ClickHouse	World's fastest analytical database	8123
PostgreSQL	Transactional database for Langfuse	5432
MongoDB	Transactional database for LibreChat	27017
MinIO	S3-compatible object storage	9090
Redis	Caching and queue	6379
Meilisearch	Full-text search for LibreChat	7700
pgvector	Vector database for RAG	5433
RAG API	Retrieval-augmented generation service for LibreChat	8001

Deploy on Railway

One-click deploy of two "Lite" variants of the stack:

• Langfuse OSS — Deploys LibreChat, the Admin Panel, Langfuse v3, and a ClickHouse MCP server pre-configured against the public ClickHouse demo cluster.
• Langfuse Cloud — Deploys LibreChat, the Admin Panel, and ClickHouse MCP, and is pre-configured to send traces to your existing Langfuse Cloud (or other remote Langfuse) project.

Both Railway templates skip Meilisearch, pgvector, and the RAG API for simplicity. The ClickHouse MCP server ships pointed at sql-clickhouse.clickhouse.com so you can chat with the public demo data immediately; the ClickHouse Cloud MCP is also wired up if you want to OAuth into your own Cloud account post-deploy. Prefer to self-host? See Quick Start below.

Quick Start

Prerequisites

• Docker and Docker Compose v2+

1. Prepare the environment

./scripts/prepare-demo.sh

This is your fastest way to get started with the Agentic Data Stack. It generates a .env file with random credentials for all services, then presents an interactive menu to optionally configure API keys for OpenAI, Anthropic, and/or Google. Any providers you skip will remain as user_provided, letting users enter their own keys in the LibreChat UI.

The script then asks whether LibreChat should send its Langfuse traces to the local Langfuse container (the default) or to a remote Langfuse project. Choose the cloud option to point LibreChat at Langfuse Cloud (or any self-hosted Langfuse) by entering the base URL, public key, and secret key for that project. You can change this later by editing LANGFUSE_BASE_URL, LANGFUSE_PUBLIC_KEY, and LANGFUSE_SECRET_KEY in .env.

You can also generate credentials separately and customize the initial administrator account credentials:

USER_EMAIL="you@example.com" USER_PASSWORD="supersecret" USER_NAME="YourName" ./scripts/generate-env.sh

Learn more about configuring your LibreChat instance at https://librechat.ai/docs.

Note: To use LibreChat's file search / RAG features, the RAG API needs a real API key for embeddings — user_provided won't work because the RAG API calls the embeddings endpoint directly. If OPENAI_API_KEY is set to user_provided, set RAG_OPENAI_API_KEY to a valid OpenAI key (it overrides OPENAI_API_KEY for RAG only). You can also switch embedding providers via EMBEDDINGS_PROVIDER (openai, azure, huggingface, huggingfacetei, ollama). See the RAG API docs for details.

2. Start the stack

docker compose up -d

3. Access the services

• LibreChat — http://localhost:3080
• Admin Panel — http://localhost:3081
• Langfuse — http://localhost:3000
• MinIO Console — http://localhost:9091 (Find credentials in .env under MINIO_ROOT_* fields)

An admin user is created automatically on first startup using the credentials from your .env file.

Architecture

LibreChat connects to ClickHouse through the MCP server, allowing AI agents to query and analyze your data. All LLM interactions are traced in Langfuse for observability, evaluation, and prompt management. The Admin Panel provides a browser-based UI for LibreChat configuration without editing librechat.yaml by hand.

Scripts

Script	Description
scripts/prepare-demo.sh	Generate .env and interactively configure API keys
scripts/generate-env.sh	Generate .env with random credentials
scripts/reset-all.sh	Stop all containers and wipe all data/volumes
scripts/create-librechat-user.sh	Manually create a LibreChat admin user
scripts/init-librechat-user.sh	Auto-init user on container startup (used internally)

Configuration

• LibreChat — librechat.yaml configures endpoints, MCP servers, and agent capabilities. The Admin Panel (http://localhost:3081) offers a browser-based alternative for most of these settings.
• Environment — .env holds all credentials and service configuration (see .env.example for reference).
• Docker — docker-compose.yml includes the four compose files:
  • langfuse-compose.yml — Langfuse, ClickHouse, PostgreSQL, Redis, MinIO
  • clickhouse-mcp-compose.yml — ClickHouse MCP server
  • librechat-compose.yml — LibreChat, MongoDB, Meilisearch, pgvector, RAG API
  • admin-panel-compose.yml — LibreChat Admin Panel

Reset Everything

To tear down all containers and delete all data:

./scripts/reset-all.sh

Then set up again and start fresh:

./scripts/prepare-demo.sh
docker compose up -d

Links

• clickhouse.ai — Project homepage
• Documentation — Full setup guide for adding ClickHouse MCP to LibreChat
• ClickHouse MCP — MCP server for ClickHouse
• LibreChat — Chat UI
• LibreChat Admin Panel — Browser-based configuration UI for LibreChat
• Langfuse — LLM observability