Set default tools

[davepgreene]

There are a few tools such as @copilot, @neovim, etc. that I want to always provide to the LLM. Looking at the config, it seems as though providing a table to config.tools like so:

{
  "CopilotC-Nvim/CopilotChat.nvim",
  build = "make tiktoken",
  dependencies = {
    {
      "MeanderingProgrammer/render-markdown.nvim",
      ft = { "markdown", "copilot-chat" },
    },
  },
  opts = {
    model = "gpt-4.1", -- AI model to use
    tools = { "@copilot", "@neovim" },
  }
}

should do it but when I use chat and ask the model to read from the workspace (what I'd expect @copilot to imply per docs) it says it can't. I thought maybe I could provide the same tools as values in the sticky table but the same issue seems to occur.

Is there a proper way to do this?

Additionally, is there a way to automatically approve specific tools provided by mcphub based on the auto-approve setting in mcphub?

[deathbeam]

For the default tools config, you need to use tool names directly, e.g without @. @ is needed only when using the tools directly in prompt.

Also do not make issue for multiple requests/questions, but for auto-approval, no it is currently not planned (i personally think auto approval is dangerous and should not be an option)

[gennaro-tedesco]

Sorry to jump on this, I too am having problems with the calling of default tools. In particular it seems the tools are "attached" but the results are not processed, see attached screenshot

as you see from the configuration, the copilot tool is defined in the options and correctly called by the prompt. However the response then stops, as the model doesn't actually process the task (to check for all such files in the codebase). Shall one re-send the prompt again afterwards (but doing so I get some strange error) or is it supposed to just work out of the box?

[davepgreene]

Yes you sould resend. Basically what is happening is that you ask LLM to do X, LLM finds appropriate tool so you get prompt for confirmation, so you have to send this prompt now (or delete the tool call from the prompt)

Is that extra step truly necessary? It seems like it's kind of defeating the point of tooling calling and agentic interaction more generally. That's kind of what I was getting at in my original secondary question about auto-approval.

In contrast, Avante forces the user to approve each tool call as it's being made unless the user has specifically approved that tool in mcphub. To me, that's a nice contrast between security and usability with an eye towards automation.

Just my 2 cents. Obviously it's y'all's project and I'm just a user. Keep up the good work! 😄

[deathbeam]

It replicates exactly how the API functions instead of hiding the implementation so yea I like how it is done. Also you can approve multiple tool calls like this at once because it combines them in 1 prompt. Its not very different from avante just asking you to confirm, but instead of some kind of confirmation it just makes prompt instead, but in the end same result, except there is no option to auto confirm basically.

[gennaro-tedesco]

Yes you sould resend.

Shall one just resend the prompt (<CR>) or shall one explicitly attach again the #glob:toolu_vrtx_...? If I just resend it seems that the model "re-calculates" the tool instead of just attaching the result (in this case the list of project files).

[deathbeam]

yea just sending the prompt with should be enoguh, as long as the tool result was success, if not then it will just try some different tool sometimes right after

[deathbeam]

for example this. it used grep to find the file (grep returns list of filenames) and then ran different tool to actually get file contents right after