Error use_mcp_tool with Avante.nvim

[alisanoelia]

╭─  use_mcp_tool  failed
│   running tool
│   Calling tool `search_repositories` on server `github` with arguments: {}
╰─  Error: missing required parameter: query

Use github mcp server.

[ravitemer]

@alisanoelia This is issue with the model not being able to call the tool with proper arguments. It happens sometimes. Can you use some other model or try submitting again so that it rectifies in the next prompt?

[alisanoelia]

@ravitemer Github MCP Server installed

• Datetime: 2025-07-09 08:23:58
• Model: gemini/gemini-2.5-flash
• Selected files:
  • /home/asha/.config/nvim/lua/plugins/mcphub.lua

Cuentame sobre el repositorio alisanoelia/st

Voy a buscar información sobre el repositorio alisanoelia/st en GitHub. Para ello, utilizaré la herramienta search_repositories.

╭─ use_mcp_tool failed
│ running tool
│ Calling tool search_repositories on server github with arguments: {}
╰─ Error: missing required parameter: query

Me disculpo, cometí un error en la llamada anterior. El parámetro query para search_repositories debe ser solo el nombre del repositorio. Intentaré de nuevo, buscando el repositorio st del propietario alisanoelia.

╭─ use_mcp_tool failed
│ running tool
│ Calling tool search_repositories on server github with arguments: {}
╰─ Error: missing required parameter: query

Me disculpo por los errores persistentes. Parece que el parámetro query de la herramienta search_repositories no está funcionando como esperaba con la sintaxis de búsqueda de GitHub.

Voy a intentar una búsqueda más simple, solo con el nombre del repositorio "st", y ver si eso funciona. Si lo hace, intentaré refinar la búsqueda para incluir el propietario alisanoelia de otra manera.

╭─ use_mcp_tool failed
│ running tool
│ Calling tool search_repositories on server github with arguments: {}
╰─ Error: missing required parameter: query

Parece que estoy teniendo problemas con el parámetro query de la herramienta search_repositories. Aunque lo estoy proporcionando, el sistema sigue indicando que falta. Esto podría deberse a un problema con la herramienta o con la forma en que se interpreta la consulta.

Antes de intentar otra búsqueda, voy a verificar el estado de los servidores MCP para asegurarme de que el servidor github está conectado y funcionando correctamente.

╭─ use_mcp_tool succeeded
│ running tool
│ Calling tool get_current_servers on server mcphub with arguments: {}
╰─ tool finished

El servidor github está conectado, por lo que el problema no es con la conexión del servidor. Parece que la herramienta search_repositories no está funcionando como se espera, ya que sigue indicando que falta el parámetro query a pesar de que lo estoy proporcionando.

Voy a intentar buscar el repositorio directamente en la web utilizando la herramienta web_search. Esto debería permitirme obtener información sobre el repositorio alisanoelia/st de una manera diferente.

╭─ web_search failed
│ running tool
│ provider: tavily
│ query: github alisanoelia/st
╰─ Error: Environment variable TAVILY_API_KEY is not set

Lo siento, parece que no puedo obtener información sobre el repositorio alisanoelia/st en este momento.

La herramienta search_repositories está dando un error inesperado con el parámetro query. Además, la herramienta web_search requiere que la variable de entorno TAVILY_API_KEY esté configurada, lo cual no es el caso.

Debido a estas limitaciones técnicas, no puedo cumplir con tu solicitud de contarte sobre el repositorio.

✓ Task Completed

No puedo proporcionarte información sobre el repositorio alisanoelia/st en este momento. La herramienta search_repositories está presentando un error con el parámetro de búsqueda, y la herramienta web_search requiere una clave de API (TAVILY_API_KEY) que no está configurada. Por lo tanto, no puedo acceder a la información del repositorio.

[ravitemer]

- Datetime: 2025-07-09 14:15:25
- Model:    copilot/gpt-4.1

> Tell me about the repository alisanoelia/st using the github server tools


I'll use the GitHub server tools to gather information about the alisanoelia/st repository, including its description, recent activity, and any other relevant details.


🤔 Thoughts

> To give a full overview, I need to retrieve the repository's metadata, description, and possibly its README or recent activity. I'll start by searching for the repository details and its README if available.


╭─  web_search  succeeded
│   running tool
│   provider: tavily
│   query: alisanoelia/st github
╰─  tool finished


🤔 Thoughts

> The initial web search confirms that alisanoelia/st is a GitHub repository focused on glyphs and icons issues, possibly related to the st terminal. To provide a detailed overview, I need to use the GitHub server tools to fetch the repository's metadata, such as its description, topics, star count, and other relevant details.


╭─  use_mcp_tool  succeeded
│   running tool
│   Calling tool `search_repositories` on server `github` with arguments: {
│     query = "alisanoelia/st"
│   }
╰─  tool finished


╭─  use_mcp_tool  failed
│   running tool
│   Calling tool `get_file_contents` on server `github` with arguments: {
│     owner = "alisanoelia",
│     path = "README.md",
│     repo = "st"
│   }
╰─  Error: Failed to get file contents. The path does not point to a file or directory, or the file does not exist in the repository.


🤔 Thoughts

> The repository metadata shows that alisanoelia/st is a public fork of the st (suckless terminal) project, written in C and under the MIT license. The description is 'MI fork st-suckless terminal.' There is no README.md in the repository. To provide additional details, I can check for topics, recent commits, or a summary of the filesystem structure.


╭─  use_mcp_tool  succeeded
│   running tool
│   Calling tool `list_branches` on server `github` with arguments: {
│     owner = "alisanoelia",
│     repo = "st"
│   }
╰─  tool finished


╭─  use_mcp_tool  succeeded
│   running tool
│   Calling tool `get_commit` on server `github` with arguments: {
│     owner = "alisanoelia",
│     repo = "st",
│     sha = "main"
│   }
╰─  tool finished


🤔 Thoughts

> I've gathered enough detail for a short report: the repository is a personal fork of the st (suckless terminal) project, written in C, under the MIT license, with no stars or forks. The latest commit (July 2025) updates color scheme files. There is no README. I'll summarize this for the user.


✓  Task Completed

The GitHub repository alisanoelia/st is a personal fork of the st (suckless terminal) project, primarily written in C and licensed under the MIT license. The repository is public, has no README file, and currently has no stars or forks. Its description is "MI fork st-suckless terminal." The most recent update (July 2025) involved changes to color scheme files (such as colorschemes/asha.h, ashes.h, grubvox.h, and pome.h), reflecting ongoing customization of the terminal's appearance. The repository does not contain a README, so documentation is minimal. For more details, visit: https://github.com/alisanoelia/st

---

@alisanoelia As you can see it varies with the model. The gemini-2.5-flash might be unable to form proper tool arguments. You can confirm this by inspecting the chat history which will be typically saved at /home/ubuntu/.local/state/nvim/avante/projects/[foldername]/history/[id].json. In the following you can see that the tool_input has query field.

Can you inspect or share your conversation history json file contents to see if its the model that is messing up the tool calls or the plugin?

[pradigunara]

@ravitemer joining the discussion as I faced the same issue. (also with gemini 2.5 pro)

From checking the history seems like gemini was passing XML format as arguments instead of json.
Also asked it which prompt that instruct it to use XML, it was pointing to this one:
https://github.com/yetone/avante.nvim/blob/c20e4e8f05b79015f1c7a2178da5298e431e58a8/lua/avante/utils/prompts.lua#L17-L27

when I specifically asked to pass json into the tool call, it succeded:

> can you try check context7 again but provide the argument with json instead of xml


That's an excellent suggestion for debugging the issue. It's possible that the MCP server expects the `tool_input` argument to be a JSON string rather than a nested XML-like structure. I will try calling the `context7` `resolve-library-id` tool again, but this time I will format the `tool_input` parameter as a JSON object, just as you suggested.

I will attempt to resolve the library ID for "counterfeiter" with this new format.


╭─  use_mcp_tool  succeeded
│   running tool
│   Calling tool `resolve-library-id` on server `context7` with arguments: {
│     libraryName = "counterfeiter"
│   }
╰─  tool finished


Excellent! It appears that your suggestion to use a JSON object for the `tool_input` was correct. The tool call succeeded this time, and I received a list of available libraries from the `context7` server.

adding custom instruction kinda solved it though

"nativeMCPServers": {
    "mcphub": {
      "custom_instructions": {
        "text": "for function calls provided by MCPHub, please use JSON as arguments instead of XML."
      }
    },

[ravitemer]

Thanks for sharing that. Can you share your avante plugin config?

[pradigunara]

sure, these are my config with lazyvim:

  {
    "yetone/avante.nvim",
    event = "VeryLazy",
    lazy = false,
    version = false, -- set this if you want to always pull the latest change
    opts = {
      -- ===== MCPHub Integration =====
      -- system_prompt as function ensures LLM always has latest MCP server state
      -- This is evaluated for every message, even in existing chats
      system_prompt = function()
        local hub = require("mcphub").get_hub_instance()
        return hub and hub:get_active_servers_prompt() or ""
      end,
      -- Using function prevents requiring mcphub before it's loaded
      custom_tools = function()
        return {
          require("mcphub.extensions.avante").mcp_tool(),
        }
      end,
      -- ==============================

      provider = "gemini",
      web_search_engine = {
        provider = "brave",
      },
      windows = {
        width = 50,
      },
      providers = {
        gemini = {
          model = "gemini-2.5-pro",
          extra_request_body = {
            temperature = 0.7,
          },
        },
      },
    },
    build = "make",
    dependencies = {
      "stevearc/dressing.nvim",
      "nvim-lua/plenary.nvim",
      "MunifTanjim/nui.nvim",
    },
  },
  {
    "ravitemer/mcphub.nvim",
    dependencies = {
      "nvim-lua/plenary.nvim",
    },
    build = "bundled_build.lua", -- Bundles `mcp-hub` binary along with the neovim plugin
    config = function()
      require("mcphub").setup({
        use_bundled_binary = true, -- Use local `mcp-hub` binary
      })
    end,
  }

[ravitemer]

@alisanoelia @pradigunara Thanks for sharing this. use_ReAct_prompt is the culprit here. For gemini adapter, avante sets this to true by default. When it is set to true, avante converts the function definition into xml format and adds it to the prompt. If we set the use_ReAct_prompt to false, everything works fine because we send the proper function tool definition. e.g

    gemini = {
      model = "gemini-2.5-flash",
      use_ReAct_prompt = false,
    }

However, some gemini models might not support function tools and for that reason avante might have set it to true by default. So, I just pushed a fix 64dccde which explictly tells in the description of the tool_input that it must be a JSON string. Seems to work fine.

You don't need to change anything in your config. Can you update the mcphub.nvim plugin and let me know if this works?

[alisanoelia]

It's work for me! 🚀 Thanks!

[pradigunara]

confirming that it also works on my side (after updating the lib and removing the custom instruction), thank you!