Model Context Protocol (STDIO)

Voxta Utilities: Model Context Protocol (STDIO) lets Voxta launch local MCP tool servers as child processes and expose their tools to the chat. Same idea as the HTTP/SSE variant — but instead of pointing at a remote URL, Voxta starts the executable itself and talks to it over stdin/stdout.

This is the variant most people want. Almost every MCP server in the public ecosystem ships as a CLI binary (Python script, npm package, Go executable) and runs locally.

What MCP gives you

MCP (Model Context Protocol) is an open standard from Anthropic for exposing tools, resources, and prompts to LLMs. Once a server is wired up, every tool it exposes becomes an action the AI can call from chat — read a file, list a directory, search the web, query a database, control an app, etc.

Common servers people install:

• @modelcontextprotocol/server-filesystem — read/write/list files in a specified folder.
• @modelcontextprotocol/server-everything — demo server that exposes every tool/resource/prompt type at once. Great for testing.
• @modelcontextprotocol/server-brave-search — web search via Brave (needs an API key).
• Plenty more at modelcontextprotocol.io.

Prerequisites

Most MCP servers in the wild are Node-based and run via npx. So in practice you'll need:

• Node.js installed on the machine running Voxta. Download from nodejs.org — the LTS build is fine.
• (Or Python + uvx / pipx if the server you want is Python-based.)

After installing Node, confirm npx --version works in a fresh terminal before configuring Voxta — if it doesn't, Voxta won't find it either.

Setup

Add the module via Manage Services → + Add Services → Voxta Utilities: Model Context Protocol (STDIO).

Open its settings and you'll find a Servers list — each row defines one MCP server, and all enabled rows run side by side. Toggling a row off keeps its config but stops the process.

Per-row fields:

Each row also has an augmentation key — that's the name your character or scenario will use to reference the server (more on this below).

Example — filesystem access

This is the "let the AI save files to my PC" recipe.

• Command: npx
• Arguments:

  -y
  @modelcontextprotocol/server-filesystem
  C:\Users\You\Documents\AI

• Working Directory: (empty)
• Environment Variables: (empty)

The last argument is the folder the AI is allowed to read and write inside. Pick a dedicated folder — the server respects it as a sandbox, so don't point it at C:\ or your whole user profile.

Example — try everything (demo)

Useful for confirming MCP is wired correctly before you commit to a specific server.

• Command: npx
• Arguments:

  -y
  @modelcontextprotocol/server-everything

Wiring it to a character or scenario

This is the step that's easy to miss. Installing the module and adding a server row is not enough on its own — the AI also needs the augmentation explicitly enabled on whatever character (or scenario) you're chatting with.

Each server row is exposed as an augmentation named mcp.<augmentation_key>. To enable it on a character:

1. Open the character you want to use it with in Studio → Characters → Edit.
2. Switch to the Configuration tab.
3. Click Add Augmentations.
4. In the popup, find the MCP server you configured (it'll be listed as mcp.<your_augmentation_key>) and select it.
5. Save the character.

Same flow exists on Scenarios if you'd rather attach the tool there.

Without this step, the AI has no idea the tools exist and will happily pretend to call them — saying things like "here's the list of files…" with nothing to back it up. If the AI claims it ran a tool but nothing actually happened, this is almost always the cause.

How to verify it's working

Three checkpoints, in order — work down the list until you find the break.

1. Did the server process start? Open the Terminal (press F2) and look for mcp-stdio log lines. If you see npx not found, command not found, or the process exits immediately, it's a launch problem (see Windows gotchas below).
2. Did the tools register? Open the Chat Inspector (Shift+Alt+I) and switch to the Functions tab. Every tool your MCP server exposes should be listed there. If the list is empty, the augmentation isn't reaching this chat — re-check the character's Configuration tab and chat style.
3. Is the model receiving them? If the tools appear in the Inspector but the AI never calls them, open the Diagnostics page and check the rendered prompt — confirm the tool definitions are actually in what gets sent to the LLM. You can also explicitly ask the AI ("use the filesystem tool to list the files in the folder") to rule out the model just choosing not to call them.

Windows gotchas

• npx not found. Child processes don't always inherit your shell's PATH. If Voxta can't find npx, put the full path in the Command field instead:

  C:\Program Files\nodejs\npx.cmd

• Folder path with spaces. Just paste the path as-is in the Arguments field — one path per line, no quoting needed. Voxta passes each line as a separate argv entry.
• First run is slow. npx -y @some/package downloads the package on first launch. Give it a minute, then it's cached.

Troubleshooting

More info

• MCP spec
• HTTP/SSE variant — for remote MCP servers.