We really need an AI model cheaper than Claude and better than Gemini
We need an Open Source model that can use MCP tools at all
Discussion
Qwen3 is by far the best for this
I did succeed in getting qwen3 to call tools, but it needs a lot of handholding to make it keep going.
How did you use it? With goose? Which model size? It normally starts to work well from 7b, and benefits a lot from a good system prompt explaining how it can interact with tools, etc. In Goose, this is managed automatically and you cannot modify the system prompt.
Do you have any good sources to start learning about MCP and that type of programming?
Easiest way to get started, IMO, would be to use VSCode with Cline or Roo, or a VSCode-derived editor like Cursor AI or Windsurf.
They all have either first-class or experimental MCP support.
Once you have it installed and connected, ask the AI to help you make a rudimentary MCP server, like a “ping” server that just logs pings in memory.
This will demonstrate how they work and how you can make your own.
If you do this, please report back on how it went!
Got a great learning experience. Got a basic MCP server that connects to my github repo to check pull requests etc.
I guess for nostr one could use it to query like content, but also provide contexts for devs, pike on how to build and integrate nostr singers or lightning, no?
Those are good ideas! Someone was putting together a Nostr Data Vending Machine (DCM) to MCP bridge. I haven’t looked into it much, but seems like a way to do it. https://dvmcp.fun/
I started working on it with cursor. Thanks for the tip, this got me going. I'll hopefully continue tomorrow. I'll keep you updated
Agree 👍
It feels like we're getting close.
Well said! 💜
Qwen3 supports tool calling
Recently did a PR to a MCP server documenting how to use it with Goose and local LLMs.
Used phi-4 model. TBH the performance isn’t that great. Not sure if it’s the MCP tools which aren’t great or the models doesn’t know what to do.
Also sent you an email a few days ago, FYI. haven’t heard back.
Have you tried Mistral AI?
Yes
In KoboldCpp we went with a different approach. Instead of relying on a models tool calling format we enforce json conformance and ask the model if it would like to use a tool. Opens up tool usage in models which are smart but lack native tool calling formats. To the user its a regular chat completions endpoint. AGPL.