Set up ClickHouse MCP server with Ollama

This guide explains how to set up Ollama with a ClickHouse MCP server.

Using ClickHouse MCP server with Ollama

This guide explains how to use the ClickHouse MCP server with Ollama.

1. Install Ollama

   Ollama is a library for running Large Language Models (LLMs) on your own machine. It has a wide range of models available and is easy to use.

   You can download Ollama for Mac, Windows, or Linux from the download page.

   Once you run Ollama, it will start a local server in the background that you can use to run models. Alternatively, you can run the server manually by running ollama serve.

   Once installed, you can pull a model down to your machine like this:

   ollama pull qwen3:8b
   

   This will pull the model to your local machine if it isn’t present. Once it’s downloaded, you can run the model like this:

   ollama run qwen3:8b
   

   Only models that have tool support will work with MCP servers.

   We can list the models that we have downloaded like this:

   ollama ls
   

   NAME                       ID              SIZE      MODIFIED
   qwen3:latest               500a1f067a9f    5.2 GB    3 days ago
   

   We can use the following command to see more information about the model that we’ve downloaded:

   ollama show qwen3
   

     Model
       architecture        qwen3
       parameters          8.2B
       context length      40960
       embedding length    4096
       quantization        Q4_K_M
   
     Capabilities
       completion
       tools
   
     Parameters
       repeat_penalty    1
       stop              "<|im_start|>"
       stop              "<|im_end|>"
       temperature       0.6
       top_k             20
       top_p             0.95
   
     License
       Apache License
       Version 2.0, January 2004
   

   We can see from this output that the default qwen3 model has just over 8 billion parameters.

2. Install MCPHost

   At the time of writing (July 2025) there is no native functionality for using Ollama with MCP servers. However, we can use MCPHost to run Ollama models with MCP servers.

   MCPHost is a Go application, so you’ll need to make sure that you have Go installed on your machine. You can then install MCPHost by running the following command:

   go install github.com/mark3labs/mcphost@latest
   

   The binary will be installed under ~/go/bin so we need to make sure that directory is on our path.

3. Configuring ClickHouse MCP server

   We can configure MCP servers with MCPHost in YAML or JSON files. MCPHost will look for config files in your home directory the following order:

   1. .mcphost.yml or .mcphost.json (preferred)
   2. .mcp.yml or .mcp.json (backwards compatibility)

   It uses a syntax that’s similar to that used in the standard MCP configuration file. Here’s an example of a ClickHouse MCP server configuration, which we’ll save to the ~/.mcphost.json file:

   {
     "mcpServers": {
       "mcp-ch": {
         "type": "local",
         "command": ["uv",
           "run",
           "--with",
           "mcp-clickhouse",
           "--python",
           "3.10",
           "mcp-clickhouse"
         ]
       }
     }
   }
   

   The main difference from the standard MCP configuration file is that we need to specify a type. The type is used to indicate the transport type used by the MCP server.

   • local → stdio transport
   • remote → streamable transport
   • builtin → inprocess transport

   We’ll also need to configure the following environment variables:

   export CLICKHOUSE_HOST=sql-clickhouse.clickhouse.com
   export CLICKHOUSE_USER=demo
   export CLICKHOUSE_PASSWORD=""
   

   In theory, you should be able to provide these variables under the environment key in the MCP configuration file, but we’ve found that this doesn’t work.

4. Running MCPHost

   Once you’ve configured the ClickHouse MCP server, you can run MCPHost by running the following command:

   mcphost --model ollama:qwen3
   

   Or, if you want to have it use a specific config file:

   mcphost --model ollama:qwen3 --config ~/.mcphost.json 
   

   If you don’t provide --model, MCPHost will look in the environment variables for ANTHROPIC_API_KEY and will use the anthropic:claude-sonnet-4-20250514 model.

   We should see the following output:

     ┃                                                                                     ┃
     ┃  Model loaded: ollama (qwen3)                                                       ┃
     ┃   MCPHost System (09:52)                                                            ┃
     ┃                                                                                     ┃
   
     ┃                                                                                     ┃
     ┃  Model loaded successfully on GPU                                                   ┃
     ┃   MCPHost System (09:52)                                                            ┃
     ┃                                                                                     ┃
   
     ┃                                                                                     ┃
     ┃  Loaded 3 tools from MCP servers                                                    ┃
     ┃   MCPHost System (09:52)                                                            ┃
     ┃                                                                                     ┃
   
     Enter your prompt (Type /help for commands, Ctrl+C to quit, ESC to cancel generation)
   

   We can use the /servers command to list the MCP servers:

     ┃                                                                                      ┃
     ┃  ## Configured MCP servers                                                           ┃
     ┃                                                                                      ┃
     ┃  1. mcp-ch                                                                           ┃
     ┃   MCPHost System (10:00)                                                             ┃
     ┃
   

   And /tools to list the tools available:

     ┃  ## Available Tools                                                                  ┃
     ┃                                                                                      ┃
     ┃  1. mcp-ch__list_databases                                                           ┃
     ┃  2. mcp-ch__list_tables                                                              ┃
     ┃  3. mcp-ch__run_select_query
   

   We can then ask the model questions about the databases/tables available in the ClickHouse SQL playground.

   In our experience when using smaller models (the default qwen3 model has 8 billion parameters), you’ll need to be more specific about what you’d like it to do. For example, you’ll need to explicitly ask it to list the databases and tables rather than straight away asking it to query a certain table. You can partially eleviate this problem by using a large model (e.g. qwen3:14b), but that will run more slowly on consumer hardware.