Skip to content

LLM Suite MCP Tools

speedyrulz edited this page Oct 28, 2026 · 1 revision

MCP Tools

MCP (Model Context Protocol) lets a model use tools: read files, search the web, query a database or reverse-engineer a binary, depending on the server. The MCP Tools page (Ctrl+3) connects RulZ to MCP servers and decides how their tools reach the model.

Security. An MCP server is a program that runs with your user account's permissions (or a remote service you give access to). Only add servers you trust. Tool results are text the model reads, so treat what comes back as untrusted, especially from servers that fetch web pages or open files you did not write.

Adding servers

Servers are configured in mcp.json, in the same mcpServers format LM Studio and Cursor use. Edit it in the panel on the right of the MCP Tools page and press Save & reconnect (or Ctrl+S). The file lives in your data folder, so you can also edit it with any editor.

Add example inserts a ready-made entry. A server is either a local program (command) or a remote one (url):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "C:\\path\\to\\folder"]
    },
    "remote": {
      "url": "https://example.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}
Field Meaning
command, args The program to start and its arguments (stdio servers).
env Extra environment variables for that program, such as API keys or paths.
cwd Working folder for the program.
url, headers A remote server using streamable HTTP, and any headers it needs.
disabled Set to true to keep an entry without connecting.

Servers started with npx, uvx and similar need Node.js or uv installed. RulZ adds the usual locations (C:\Program Files\nodejs, %APPDATA%\npm, ~\.local\bin, ~\.cargo\bin) to the PATH it gives them, since an app started from Explorer doesn't always see them.

Each server card shows its state (connecting, connected, error or disabled), the tools it offers and a Log with the server's own error output. Use the reconnect button after fixing a problem.

Built-in file tools

Besides your own servers, the page always shows a built-in server called rulz with two read-only tools:

Tool What it does
list_directory Lists a folder's files and subfolders with sizes, up to three levels deep.
find_files Finds files or folders by name under a folder, using a wildcard such as *.exe or plain text.

They let the model find exact paths before it uses other tools. For example, given only C:\Program Files\Some App, it can look inside and pass the program's .exe to REA instead of guessing the name. They show names and sizes only: they can't read or change files.

They work anywhere your Windows account can read, take absolute paths (%VARIABLES% such as %ProgramData% are expanded), and ask for approval like any other tool. In on-demand mode they stay loaded, since they are small. Switch them off on their card if you don't want them.

Approving tool calls

By default, every tool call asks first. The dialog shows the arguments and offers:

  • Deny — the model is told you declined.
  • Allow once.
  • Always allow this tool — skips the question for that tool from now on.

Tools you always allow are listed on the MCP Tools page; click one to require approval again. The switch Ask before running a tool turns approval off entirely, which is only a good idea for servers and tools you fully trust. Stopping a reply (Esc) during the tool phase skips the remaining calls.

Calls and results appear in the chat as expandable blocks with the arguments and the full result.

Choosing which tools the model sees

Each tool has an on/off switch, and each server has Enable all and Disable all buttons (plus a filter box when it offers more than eight tools). Every enabled tool's description is normally sent to the model with every message, so many tools can use a lot of context. The page shows the estimated cost per server and in total.

Tool loading modes

Tool loading (on the MCP Tools page) has three settings:

  • All enabled tools — every enabled tool is described to the model in every message.
  • On demand — the model gets only a short catalog of tool names plus three built-in tools: rulz_search_tools, rulz_load_tools and rulz_unload_tools. It searches for what it needs and loads those tools; only loaded tools' full descriptions are sent.
  • Auto (default) — uses all while the enabled tools fit the tool budget (default 6,000 tokens), and on demand beyond it.

In on-demand mode the loaded tools are kept per chat and shown in the chat sidebar, where you can unload them. If loading more would exceed the budget, the least recently used tools are unloaded automatically and the model is told which. Calling a tool also keeps it loaded.

This is what makes very large servers practical on a small context window. For example, the REA reverse-engineering server offers 207 tools; sending all their descriptions takes over 200,000 tokens, but in on-demand mode a model on a 16K context can load just the four it needs for a task.

If the enabled tools alone would take more than 60% of the model's context, the chat stops with a message telling you to switch to on-demand loading, disable some tools, or load the model with a larger context.

Shortening old tool results

Tool results can be large. With Shorten older tool results (on by default), once you send a new message, results from earlier turns are sent to the model as just their beginning and end, with a note saying how much was left out and that it can run the tool again. The chat itself still shows every result in full.

Keep recent results in full (default 3) sets how many of the latest results in the turn still in progress stay complete, so a multi-step task isn't cut short. Results under 800 characters are never shortened.

Limits

  • A single tool result is capped at 24,000 characters.
  • A tool call times out after two minutes.
  • The model can make at most 12 rounds of tool calls in one reply before RulZ stops and tells you.

Turning tools off

The Use MCP tools in chats switch on the MCP Tools page (and the Tools (MCP) section in each chat's settings) stops all tools from being offered, without removing your servers.

RulZ AI Tools Suite

The apps

RulzTrainer

README (full reference) Releases

Rulz Training Captioner

Getting started

Using the app

Models and GPUs

Reference

GitHub repository

RulZ LLM Suite

Using the app

Reference

Contributing


Releases · Issues · Source

RulZ Coding Agent

Using the app

Reference

Contributing


Releases · Issues · Source

RulZ Image Organizer

Using RulZ

More

RulZ Image Tournament

Using the app

Developers

RulZ Image Crawler

Getting started

Using the app

Help

README (full reference)

RulZ Image Preparer

Getting started

Using the app

Reference

README (full reference)

RulZ Image Grid

Using the app

Help

README (full reference)

RulZ Simple Video Clip Extractor

Getting started

User guide

Help

Clone this wiki locally