DOC: Installing MCP-Bridge with Ollama in Open-WebUI

[rgaricano]

Thanks man!!!
(my penny)

Installing MPC-Bridge with Ollama in Open-WebUI

Clone repo
git clone https://github.com/SecretiveShell/MCP-Bridge.git

Install MPC-bridge,

cd MCP-Bridge
uv sync

make your mpc-bridge config file, create config.json
(models are not necessary, it take its as other connections)
e.g. (inference_server is ollama's installed network address, network is config network address of MCP_Bridge server)

{
   "inference_server": {
      "base_url": "http://127.0.0.1:11434/v1",
      "api_key": "None"
   },
    "mcp_servers": {
        "fetch": {
            "command": "uvx",
            "args": [
                "mcp-server-fetch"
            ]
        }
    },
    "sampling": {
        "timeout": 10,
        "models": [
            {
                "model": "llama3.2:latest",
                "intelligence": 0.5,
                "cost": 0,
                "speed": 0.6
            },
            {
                "model": "deepseek-r1:8b",
                "intelligence": 0.7,
                "cost": 0,
                "speed": 0.4
            }
        ]
    },
    "network": {
        "host": "0.0.0.0",
        "port": 9090
    },
    "logging": {
        "log_level": "DEBUG"
    }
}

run server
uv run mcp_bridge/main.py
& check thats is runnig

On openwebui add a openai api compatible connection for mpc-bridge server, and check:

Now you have all models as normal, but duplicated (one from ollama and one from mpc-bridge, one with size other without), to no be confused you can switch off ollama connection.

And voila, you have all models on tab & you can chat telling for use tools, or ask for avaible tools ;-)

(if is running fine you can change log_level to "INFO", install it as service,...)

open-webui/open-webui#7363 (comment)

[SecretiveShell]

This is a great resource!

I would love to see it in the official documentation, if you are happy to submit a pr.

[gabrieligbastos]

Awesome, great resource! I followed it and now able to talk it through bridge with openwebui.
The problem im getting, it that is kind of like this =/

What did i miss ?
I did:

ollama pull llama3.2
ollama run llama3.2

and all other steps of your tutorial :)

[SecretiveShell]

This is caused by the tool call parser inside of ollama failing to recognise the start of a tool call and putting the tokens into the chat completion.

I would suggest using a different model because llama3.2 1b/3b are incredibly weak when it comes to tool calling. Look around 32b parameter size and make sure the template has tool calling support. Llama 3.1 8b is around the minimum bar from my experience.

[gabrieligbastos]

Awesome, the reason was a mvp to just make it work. Did ran with qwq 32b model to test and worked! thanks :)