|
| 1 | +--- |
| 2 | +slug: /use-cases/AI/MCP/ollama |
| 3 | +sidebar_label: 'Integrate Ollama' |
| 4 | +title: 'Set Up ClickHouse MCP Server with Ollama' |
| 5 | +pagination_prev: null |
| 6 | +pagination_next: null |
| 7 | +description: 'This guide explains how to set up Ollama with a ClickHouse MCP server.' |
| 8 | +keywords: ['AI', 'Ollama', 'MCP'] |
| 9 | +show_related_blogs: true |
| 10 | +--- |
| 11 | + |
| 12 | +import {CardHorizontal} from '@clickhouse/click-ui/bundled' |
| 13 | +import Link from '@docusaurus/Link'; |
| 14 | +import Image from '@theme/IdealImage'; |
| 15 | + |
| 16 | +# Using ClickHouse MCP server with Ollama |
| 17 | + |
| 18 | +> This guide explains how to use the ClickHouse MCP Server with Ollama. |
| 19 | +
|
| 20 | +<VerticalStepper headerLevel="h2"> |
| 21 | + |
| 22 | +## Install Ollama {#install-ollama} |
| 23 | + |
| 24 | +Ollama is a library for running Large Language Models (LLMs) on your own machine. |
| 25 | +It has a [wide range of models available](https://ollama.com/library) and is easy to use. |
| 26 | + |
| 27 | +You can download Ollama for Mac, Windows, or Linux from the [download page](https://ollama.com/download). |
| 28 | + |
| 29 | +Once you run Ollama, it will start a local server in the background that you can use to run models. |
| 30 | +Alternatively, you can run the server manually by running `ollama serve`. |
| 31 | + |
| 32 | +Once installed, you can pull a model down to your machine like this: |
| 33 | + |
| 34 | +```bash |
| 35 | +ollama pull qwen3:8b |
| 36 | +``` |
| 37 | + |
| 38 | +This will pull the model to your local machine if it is not present. |
| 39 | +Once it's downloaded, you can run the model like this: |
| 40 | + |
| 41 | +```bash |
| 42 | +ollama run qwen3:8b |
| 43 | +``` |
| 44 | + |
| 45 | +:::note |
| 46 | +Only [models that have tool support](https://ollama.com/search?c=tools) will work with MCP Servers. |
| 47 | +::: |
| 48 | + |
| 49 | +We can list the models that we have downloaded like this: |
| 50 | + |
| 51 | +```bash |
| 52 | +ollama ls |
| 53 | +``` |
| 54 | + |
| 55 | +```text |
| 56 | +NAME ID SIZE MODIFIED |
| 57 | +qwen3:latest 500a1f067a9f 5.2 GB 3 days ago |
| 58 | +``` |
| 59 | + |
| 60 | +We can use the following command to see more information about the model that we've downloaded: |
| 61 | + |
| 62 | +```bash |
| 63 | +ollama show qwen3 |
| 64 | +``` |
| 65 | + |
| 66 | +```text |
| 67 | + Model |
| 68 | + architecture qwen3 |
| 69 | + parameters 8.2B |
| 70 | + context length 40960 |
| 71 | + embedding length 4096 |
| 72 | + quantization Q4_K_M |
| 73 | +
|
| 74 | + Capabilities |
| 75 | + completion |
| 76 | + tools |
| 77 | +
|
| 78 | + Parameters |
| 79 | + repeat_penalty 1 |
| 80 | + stop "<|im_start|>" |
| 81 | + stop "<|im_end|>" |
| 82 | + temperature 0.6 |
| 83 | + top_k 20 |
| 84 | + top_p 0.95 |
| 85 | +
|
| 86 | + License |
| 87 | + Apache License |
| 88 | + Version 2.0, January 2004 |
| 89 | +``` |
| 90 | + |
| 91 | +We can see from this output that the default qwen3 model has just over 8 billion parameters. |
| 92 | + |
| 93 | +## Install MCPHost {#install-mcphost} |
| 94 | + |
| 95 | +At the time of writing (July 2025) there is no native functionality for using Ollama with MCP Servers. |
| 96 | +However, we can use [MCPHost](https://github.com/mark3labs/mcphost) to run Ollama models with MCP Servers. |
| 97 | + |
| 98 | +MCPHost is a Go application, so you'll need to make sure that you have [Go installed](https://go.dev/doc/install) on your machine. |
| 99 | +You can then install MCPHost by running the following command: |
| 100 | + |
| 101 | +```bash |
| 102 | +go install github.com/mark3labs/mcphost@latest |
| 103 | +``` |
| 104 | + |
| 105 | +The binary will be installed under `~/go/bin` so we need to make sure that directory is on our path. |
| 106 | + |
| 107 | +## Configuring ClickHouse MCP Server {#configure-clickhouse-mcp-server} |
| 108 | + |
| 109 | +We can configure MCP Servers with MCPHost in YAML or JSON files. |
| 110 | +MCPHost will look for config files in your home directory the following order: |
| 111 | + |
| 112 | +1. `.mcphost.yml` or `.mcphost.json` (preferred) |
| 113 | +2. `.mcp.yml` or `.mcp.json` (backwards compatibility) |
| 114 | + |
| 115 | +It uses a syntax that's similar to that used in the standard MCP configuration file. |
| 116 | +Here's an example of a ClickHouse MCP server configuration, which we'll save to the `~/.mcphost.json` file: |
| 117 | + |
| 118 | +```json |
| 119 | +{ |
| 120 | + "mcpServers": { |
| 121 | + "mcp-ch": { |
| 122 | + "type": "local", |
| 123 | + "command": ["uv", |
| 124 | + "run", |
| 125 | + "--with", |
| 126 | + "mcp-clickhouse", |
| 127 | + "--python", |
| 128 | + "3.10", |
| 129 | + "mcp-clickhouse" |
| 130 | + ] |
| 131 | + } |
| 132 | + } |
| 133 | +} |
| 134 | +``` |
| 135 | + |
| 136 | +The main difference from the standard MCP configuration file is that we need to specify a `type`. |
| 137 | +The type is used to indicate the transport type used by the MCP Server. |
| 138 | + |
| 139 | +* `local` → stdio transport |
| 140 | +* `remote` → streamable transport |
| 141 | +* `builtin` → inprocess transport |
| 142 | + |
| 143 | +We'll also need to configure the following environment variables: |
| 144 | + |
| 145 | +```bash |
| 146 | +export CLICKHOUSE_HOST=sql-clickhouse.clickhouse.com |
| 147 | +export CLICKHOUSE_USER=demo |
| 148 | +export CLICKHOUSE_PASSWORD="" |
| 149 | +``` |
| 150 | + |
| 151 | +:::note |
| 152 | +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. |
| 153 | +::: |
| 154 | + |
| 155 | +## Running MCPHost {#running-mcphost} |
| 156 | + |
| 157 | +Once you've configured the ClickHouse MCP server, you can run MCPHost by running the following command: |
| 158 | + |
| 159 | +```bash |
| 160 | +mcphost --model ollama:qwen3 |
| 161 | +``` |
| 162 | + |
| 163 | +Or, if you want to have it use a specific config file: |
| 164 | + |
| 165 | +```bash |
| 166 | +mcphost --model ollama:qwen3 --config ~/.mcphost.json |
| 167 | +``` |
| 168 | + |
| 169 | +:::warning |
| 170 | +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. |
| 171 | +::: |
| 172 | + |
| 173 | +We should see the following output: |
| 174 | + |
| 175 | +```text |
| 176 | + ┃ ┃ |
| 177 | + ┃ Model loaded: ollama (qwen3) ┃ |
| 178 | + ┃ MCPHost System (09:52) ┃ |
| 179 | + ┃ ┃ |
| 180 | +
|
| 181 | + ┃ ┃ |
| 182 | + ┃ Model loaded successfully on GPU ┃ |
| 183 | + ┃ MCPHost System (09:52) ┃ |
| 184 | + ┃ ┃ |
| 185 | +
|
| 186 | + ┃ ┃ |
| 187 | + ┃ Loaded 3 tools from MCP servers ┃ |
| 188 | + ┃ MCPHost System (09:52) ┃ |
| 189 | + ┃ ┃ |
| 190 | +
|
| 191 | + Enter your prompt (Type /help for commands, Ctrl+C to quit, ESC to cancel generation) |
| 192 | +``` |
| 193 | + |
| 194 | +We can use the `/servers` command to list the MCP Servers: |
| 195 | + |
| 196 | +```text |
| 197 | + ┃ ┃ |
| 198 | + ┃ ## Configured MCP Servers ┃ |
| 199 | + ┃ ┃ |
| 200 | + ┃ 1. mcp-ch ┃ |
| 201 | + ┃ MCPHost System (10:00) ┃ |
| 202 | + ┃ |
| 203 | +``` |
| 204 | + |
| 205 | +And `/tools` to list the tools available: |
| 206 | + |
| 207 | +```text |
| 208 | + ┃ ## Available Tools ┃ |
| 209 | + ┃ ┃ |
| 210 | + ┃ 1. mcp-ch__list_databases ┃ |
| 211 | + ┃ 2. mcp-ch__list_tables ┃ |
| 212 | + ┃ 3. mcp-ch__run_select_query |
| 213 | +``` |
| 214 | + |
| 215 | +We can then ask the model questions about the databases/tables available in the ClickHouse SQL playground. |
| 216 | + |
| 217 | +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. |
| 218 | +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. |
| 219 | +You can partially eleviate this problem by using a large model (e.g. qwen3:14b), but that will run more slowly on consumer hardware. |
| 220 | + |
| 221 | +</VerticalStepper> |
0 commit comments