MCP Resources -> Chat Variables

[ravitemer]

Hi guys, with v4 of mcphub.nvim

• You can use MCP resources as chat #variables. This gives you more control while giving more context and also saves an API call from the LLM trying to access the resource. Now, with Lua Native MCP Servers you can create your own custom resources, tools, templates easily.

When make_vars = true:

• MCP resources appear as #variables in chat
• Variables update in real-time through chat.variables table

• E.g LSP current file diagnostics

• Can configure how the result of the mcp tool calls are handled. By default the result is added to chat history. If you wish to see the result of mcp tool call, set show_result_in_chat to true. (Note: if the result text is markdown with headers, text after the headers wont be picked up by codecompanion)

1. MCPHub Configuration:

require("mcphub").setup({
    extensions = {
        codecompanion = {
            -- Show the mcp tool result in the chat buffer
            show_result_in_chat = true,
            -- Make chat #variables from MCP server resources
            make_vars = true,
        }
    }
})

2. CodeCompanion Configuration:

require("codecompanion").setup({
    strategies = {
        chat = {
            tools = {
                ["mcp"] = {
                    -- Prevent mcphub from loading before needed
                    callback = function() 
                        return require("mcphub.extensions.codecompanion") 
                    end,
                    description = "Call tools and resources from the MCP Servers",
                    opts = {
                        requires_approval = true,
                    }
                }
            }
        }
    }
})

Tool Execution

• Use vim.g.codecompanion_auto_tool_mode = true or gta in chat for auto-approval
• Tools from all enabled MCP servers are available
• Resources can be accessed via their URIs

Note: CodeCompanion doesn't reevaluate tools already added to a chat. So, any servers disabled or enabled doesn't reflect in the system prompt unless you create a new chat. However, #variables are always kept up-to-date by updating config.strategies.chat.variables which is used by cmp.

For more information visit:
https://github.com/ravitemer/mcphub.nvim/wiki/CodeCompanion

[ravitemer]

Guys, Now mcphub manages auto approval behavior of tool calls rather than depending on codecompanion. This also respects the gta or codecompanion_auto_tool_mode though.

We can toggle the auto approval for our mcphub with ga in the Hub view. This internally toggles the vim.g.mcphub_auto_approve variable.

With this update many issues with auto approvals are fixed. We right away return error if the tool was called with incorrect parameters rather than asking for an approval first. We can now control how we want to show the approval window (which current uses the vim.ui.confirm but a custom window might be good for lengthy text in approvals).

Changes needed

Codecompanion

Remove requires_approval from opts

require("codecompanion").setup({
    strategies = {
        chat = {
            tools = {
                ["mcp"] = {
                    -- calling it in a function would prevent mcphub from being loaded before it's needed
                    callback = function() return require("mcphub.extensions.codecompanion") end,
                    description = "Call tools and resources from the MCP Servers",
                }
            }
        }
    }
})

[olimorris]

This is incredible work! I can't wait to dive more into it once I merge the function calling PR.

Also, just loving the fact you have Todoist, Figma and Discord all within Neovim 🚀.

[hampgoodwin]

This is really exciting

Is it possible to have the results of the mcp tool also shared in the prompt? I can look at the stdio or docker logs, but that seems a bit cumbersome. Especially in the case of an error. For ex with filesystem  mcp server, there's no feedback that it was a failure.

2025-04-10 13:27:25 {"result":{"content":[{"type":"text","text":"Error: Access denied - path outside allowed directories: /projects/readme.md not in /projects/organization"}],"isError":true},"jsonrpc":"2.0","id":7}

from

<tools>
  <tool name="mcp">
    <action type="use_mcp_tool">
      <server_name><![CDATA[filesystem]]></server_name>
      <tool_name><![CDATA[write_file]]></tool_name>
      <arguments><![CDATA[{"path": "/projects/readme.md", "content": ""}]]></arguments>
    </action>
  </tool>
</tools>

The error is shared with the llm, but the prompter, doesn't see it. So I basically don't know that the operation failed without checking docker logs.

[ravitemer]

Is it possible to have the results of the mcp tool also shared in the prompt?

Sorry, what do you mean? When the tool runs the output of the tool is added to the chat (with show_result_in_chat is true) else it will be added to the chat history.