Running multiple tools with Ollamasharp

[Rpgdudester]

Good morning,

First off, I want to say that OllamaSharp is awesome,

it's been great being able to create my own MCP servers and client being able to run tools for my EVA I'm creating for internal use.

I've been able to load up my tools servers that use the standard [MCPToolType] in console apps that I've coded in c# 10 RC1, and everything's been working great with model (in this case Qwen3:8B) being able to find them and run them without an issue.

Now I'm probably going to make myself look stupid on this but how can I get the model to chain tools, one after the other to run multiple tools in order using the tool console? I can't tell if I'm doing something wrong or if the tool console was just an example/demo on how everything works together.

This is what happens when I ask it to get the last 5 tickets for the youtrack project GrowSmart Helpdesk. You can see that it walks through and plans everything out, but it's like it's not realizing that it needs to run multiple tools to get the results and just .. dies so to speak. I can't tell if it's something silly I've done or not.

Thank for creating an awesome library for us!

Joshua.

[awaescher]

This question is absolutely legit and thanks for the kind words.

Let me begin this way: The primary goal of OllamaSharp is to cover 100% of the Ollama API, providing a typesafe and easy-to-use layer to work with that API.

There's the Chat class which I guess you are using here. I would like to emphasize that this is just one way how a chat with an AI assistant can be implemented. It does not have to be this way. It's a little abstraction to simplify building chats but might not cover every use case.

The way this chat class is implemented today is that it automatically executes tools the AI model wants to be executed in a batch and then feeds back the results to the model which has the chance to respond with the information that resulted from the tool call(s). It is important to note here that the model has no chance to further iterate, as the execution ends as soon as the model is done with its message.

Your case reads a bit more agentic where the model should reflect on the tool results and whether or not it should continue to iterate (maybe calling other tools or changing the arguments and retry tool calls) until it is satisfied with the results and feels confident enough to stop iterating and write a response.

This is more than the current Chat class does. Just a little bit though. But I think that it would be too much to implement in the Chat class because others might want to handle this iteration logic by themselves and by their own rules for stopping, etc.

I would encourage you to build this with a pretty simple loop combined with a system prompt that tells the model it should signal some kind of stop word that you can react in code and to stop the iteration.

[Rpgdudester]

See, I had noticed on this page:

#218

that when you ran your test, it looked like it had called all 3 tools with no problem, which is what had me utterly confused when I was researching this over the past few days. I /almost/ went as far as cutting everything out and recoding everything to use the AI Extension console but the AI extension one wouldn't be able to load up the tools using ollamasharp's GetFromMcpServers because the AI extension one is wanting AITools vs McpClientTools and there would be no way to convert it.

I like the fact that the Ollamasharp system has been great and easy to use and set up (I even built my own MCP console and a persona manager for my models, along with TTS using Kokorosharp), this one thing has been the head scratcher for the past day or so!

At least I know I'm not going crazy so thank you for that!

Joshua.

[awaescher]

Okay wait, not that we are talking about different things. What I implemented with #218 was what I mentioned with "executed in a batch". I meant that all tool calls are executed and the results are fed to the model afterwards. This works out of the box since then if your model is smart enough to come up with multiple tool calls at once.

What I referred to in my first response was that the model got no chance to iterate on the results later on. So it cannot read the results from tool call 1 and try another tool call based on that result. I thought you wanted to achieve this as you wrote "chain tools" and "one after the other". This is something you should implement by yourself as the way you handle the responses are not generic enough to code it into the Chat class (imo).

[Rpgdudester]

Sorry for not responding earlier, I has been working on this with GPT 4o over thursday and friday then decided to step away from my PC over the weekend so I think this through.

You're correct in the fact that the chaining of tools (calling one after the other) is definitely what my intention is.

I understand that your example was showing it calling two different tools that have no bearing on each other, (population and gps coords) which does make sense. but

I mean, it works great calling a single tool with no problem, and I was able to get it to run through multiple tools within a loop but then (which could totally be the way I had my Qwen3 prompting for this) it wouldn't put a stop word like or at the end of the response when it was finished.

I'm going to keep toying with this but I appreciate you pointing me in the right direction, thanks so much!

Joshua.

[awaescher]

It's fun seeing your response now as I just finished working and pushing this which allows sequential tool calling. It allows not only having multiple tool calls in a single message but also one tool call after another in multiple messages, if the model wants to do so.

Try updating to 5.4.6, chances are this is already implementing what you need. With this, you should be able to remove the necessity for a stop word as OllamaSharp runs an internal recursion on the messages and will automatically stop the moment the AI model does not request any more tool calls.

[awaescher]

It's pretty simple to test in the tool console demo with qwen3:4b-instruct which works pretty reliably in my case.

Just use a system prompt like this:

Use tools whenever possible, but only one at a time. If you need multiple tools, call them one after another.

[Rpgdudester]

DUDE! you're freakin awesome! haha. I'll update right now and see what happens! :) thank you so much!

[Rpgdudester]

I just tested it and it definitely ran one tool right after another and didn't stop this time.

So thank you so much for doing that!

Joshua

[awaescher]

Nice, was hoping to read that. Have fun building, Joshua.