Core Tools

These tools defined in the core_tools extension provide essential functionality for LLMs to interact with their environment, perform calculations, and manage persistent data.

See Tools Docs for details on registering and using your own tools.

Math & Logic

• calc(expression: str) -> str - Evaluate a mathematical expression. Supports arithmetic, comparison, boolean operators, and common math functions.
  • expression: The mathematical expression string (e.g., sqrt(16) + 5).

Utilities

• get_current_time(tz_name: str = None) -> str - Get the current time in ISO-8601 format.
  • tz_name: Optional timezone name (e.g., 'America/New_York'). Defaults to UTC.

File System Tools

The built-in computer extension filesystem.py tools provide a native Python implementation of Anthropic's node.js Filesystem MCP Server tools:

• read_text_file - Read complete contents of a file as text
• read_media_file - Read an image or audio file
• read_multiple_files - Read multiple files simultaneously
• write_file - Create new file or overwrite existing (exercise caution with this)
• edit_file - Make selective edits using advanced pattern matching and formatting
• create_directory - Create new directory or ensure it exists
• list_directory - List directory contents with [FILE] or [DIR] prefixes
• list_directory_with_sizes - List directory contents with [FILE] or [DIR] prefixes, including file sizes
• move_file - Move or rename files and directories
• search_files - Recursively search for files/directories that match or do not match patterns
• directory_tree - Get recursive JSON tree structure of directory contents
• get_file_info - Get detailed file/directory metadata
• list_allowed_directories - List all directories the server is allowed to access

Computer Use

The built-in computer extension transforms AI agents into autonomous computer operators. Based on Anthropic's computer use tools, it enables agents to see your screen, control the mouse and keyboard, execute shell commands, and edit files - just like a human sitting at the computer.

This unlocks powerful capabilities that traditional API-based tools cannot achieve:

• Visual Verification: Confirm that code actually renders correctly in a browser
• Desktop Automation: Control any GUI application - web browsers, IDEs, terminals
• End-to-End Workflows: Chain together multiple applications in a single task
• Legacy Applications: Automate software that lacks APIs

For example, an agent can write a web application, open a browser, and capture a screenshot to prove it works:

See the Computer Use docs for complete usage details.

Code Execution Tools

LLMS includes a suite of tools for executing code in various languages within a sandboxed environment. These tools are designed to allow the agent to run scripts, perform calculations, and verify logic safely.

Supported Languages

The following tools are available:

• run_python(code) - Executes Python code.
• run_javascript(code) - Executes JavaScript code (uses bun or node).
• run_typescript(code) - Executes TypeScript code (uses bun or node).
• run_csharp(code) - Executes C# code (uses dotnet run with .NET 10+ single-file support).

These tools allow LLMs to run code snippets dynamically as part of their reasoning process.

Python and C# Examples

Run Python

Click to view full size

Run C#

Click to view full size

JavaScript and TypeScript Examples

Run JavaScript

Click to view full size

Run TypeScript

Click to view full size

Sandbox Environment

Code execution happens in a temporary, restricted environment to ensure safety and stability.

Resource Limits

To prevent infinite loops or excessive resource consumption, the following limits are enforced using ulimit (on Linux/macOS):

• Execution Time: The process is limited to 5 seconds of CPU time.
• Memory: The process is limited to 8 GB virtual memory limit.

File System

• Each execution runs in a clean, temporary directory.
• Converting LLMS_RUN_AS will try to chmod 777 this directory so the target user can write to it.
• Note: Files created during execution are transient and lost after the tool completes, unless their content is printed to stdout.

Environment Variables

• The environment is cleared to prevent leakage of sensitive tokens or API keys.
• Only the PATH variable is preserved to ensure standard tools function correctly.
• For C# (dotnet), HOME and DOTNET_CLI_HOME are set to the temporary directory to allow write access for the runtime intermediate files.

Running with Lower Privileges (LLMS_RUN_AS)

By default, code runs as the user running the LLMS process. For enhanced security, you can configure the tools to execute code as a restricted user (e.g., nobody or a dedicated restricted user).

Configuration

Set the LLMS_RUN_AS environment variable to the username you want code to run as.

Example:

export LLMS_RUN_AS=nobody

Requirements for LLMS_RUN_AS

1. Sudo Access: The user running the LLMS process must have sudo privileges configured to run commands as the target user without a password.
2. Permissions: The LLMS process attempts to grant read/write access to the temporary execution directory for the target user (via chmod 777).

Example sudoers configuration:

# Allow 'mythz' to run commands as 'nobody' without password
mythz ALL=(nobody) NOPASSWD: ALL

How it Works

When LLMS_RUN_AS is set:

1. A temporary directory is created.
2. Permissions on the directory are relaxed so the restricted user can access it.
3. The command is executed wrapped in sudo -u <user>.
   • Example: sudo -u nobody bash -c 'ulimit -t 5; ... python script.py'

This execution model ensures that even if malicious code escapes the ulimit sandbox or accesses the filesystem, it is restricted by the OS-level permissions of the nobody (or specified) user.