Wolfram/AgentTools

A Wolfram Language toolkit for integrating with AI agents and LLMs — providing MCP servers, agent skills, and other standard interfaces that give AI systems access to Wolfram's computational capabilities.

Table of Contents

• Features
• Requirements
• Installation
• Quick Start
• Predefined Servers
• Supported Clients
• Available Tools
• Creating Custom Servers
• API Reference
• Development
• License

Features

• Predefined servers for common use cases (general computation, Wolfram|Alpha queries, development)
• Semantic search across Wolfram documentation and Wolfram|Alpha results
• Code evaluation with Wolfram Language directly in AI conversations
• Notebook support for reading and writing Wolfram notebooks
• Custom servers with tailored tools for specific needs
• Wide client support including Claude Desktop, Cursor, VS Code, and more
• MCP prompts for enhanced context and workflows
• MCP Apps for interactive UI resources in supported clients (e.g., embedded notebook viewers, Wolfram|Alpha result displays)
• MCP Roots support so the server, evaluator, and external tools (TestReport and similar) resolve relative paths against the client's project directory
• Agent Skills for distributing Wolfram tools as portable skills to AI coding agents (Claude Code, Cursor, Gemini CLI, VS Code, and more)
• Paclet extensions allowing third-party paclets to contribute MCP tools, prompts, and servers via the "AgentTools" extension
• Docker image for running the MCP server in a container without a local Wolfram Engine installation

Requirements

• Wolfram Language 14.3 or higher
• An MCP-compatible client application (see Supported Clients)
• Optional: LLMKit subscription for enhanced semantic search capabilities

Installation

Install the Paclet

PacletInstall["Wolfram/AgentTools"]

Load the Package

Needs["Wolfram`AgentTools`"]

Docker Image

A prebuilt Docker image is available as an alternative to a local Wolfram Engine installation. See Docker documentation for setup and MCP client configuration examples.

docker pull ghcr.io/wolframresearch/mcpserver:latest

Quick Start

Install a Wolfram MCP server for Claude Desktop:

InstallMCPServer["ClaudeDesktop"]
(* Out: Success["InstallMCPServer", <|...|>] *)

When no server is specified, each named client gets a sensible default — chat clients (Claude Desktop, Goose) get the general-purpose "Wolfram" server, while coding clients (Claude Code, Cursor, VS Code, etc.) get "WolframLanguage". For File[...] targets, the same per-client default applies when the path or content identifies a known client (or "ApplicationName" is supplied); otherwise it falls back to "Wolfram". Pass an explicit server name as the second argument to override.

After restarting Claude Desktop, it will have access to Wolfram knowledge and tools:

To install a specific server type:

InstallMCPServer["ClaudeDesktop", "WolframLanguage"]

To uninstall:

UninstallMCPServer["ClaudeDesktop"]              (* Remove all servers *)
UninstallMCPServer["ClaudeDesktop", "Wolfram"]   (* Remove specific server *)

Predefined Servers

AgentTools includes four predefined server configurations, each optimized for different use cases:

Server	Best For	Tools
Wolfram (default)	General-purpose use combining computational power with natural language	WolframContext, WolframLanguageEvaluator, WolframAlpha
WolframAlpha	Natural language queries without code execution	WolframAlphaContext*, WolframAlpha
WolframLanguage	Wolfram Language development and learning	WolframLanguageContext, WolframLanguageEvaluator, ReadNotebook, WriteNotebook, SymbolDefinition, CodeInspector, TestReport
WolframPacletDevelopment	Developing and maintaining Wolfram paclets	All WolframLanguage tools plus documentation tools (CreateSymbolDoc, EditSymbolDoc, EditSymbolDocExamples) and paclet release tools (CheckPaclet, BuildPaclet, SubmitPaclet)

*Requires LLMKit subscription

Install a specific server:

InstallMCPServer["ClaudeDesktop", "WolframLanguage"]

See docs/servers.md for detailed information about each server and guidance on choosing the right one.

Supported Clients

AgentTools can be installed into the following MCP client applications:

Client	Name	Project Support
Amazon Q Developer	"AmazonQ"	Yes
Augment Code	"AugmentCode"	No
Augment Code IDE	"AugmentCodeIDE"	No
Claude Code	"ClaudeCode"	Yes
Claude Desktop	"ClaudeDesktop"	No
Cline	"Cline"	No
Copilot CLI	"CopilotCLI"	No
Cursor	"Cursor"	No
Gemini CLI	"GeminiCLI"	No
Goose	"Goose"	No
Google Antigravity	"Antigravity"	No
Junie (JetBrains IDE plugin + CLI)	"Junie"	Yes
OpenAI Codex	"Codex"	Yes
OpenCode	"OpenCode"	Yes
Visual Studio Code	"VisualStudioCode"	Yes
Windsurf	"Windsurf"	No
Zed	"Zed"	Yes

Project-Level Installation

Clients with project support can have servers installed for specific projects:

InstallMCPServer[{"ClaudeCode", "/path/to/project"}, "WolframLanguage"]

Claude Desktop

Claude Desktop offers an excellent integration experience with AgentTools, providing seamless access to Wolfram Language's computational capabilities.

Cursor

Install an MCP server for use in Cursor:

InstallMCPServer["Cursor"]
(* Out: Success["InstallMCPServer", <|...|>] *)

Check the MCP tab in Cursor settings to verify the server connection:

Your Wolfram tools will now be available in Cursor agent chat:

Other Clients

AgentTools works with any stdio-based MCP client. See docs/mcp-clients.md for manual configuration instructions.

Available Tools

AgentTools provides a variety of tools organized by category:

Context Tools (Semantic Search)

Search Wolfram resources using semantic similarity:

• WolframContext - Combines the functionality of WolframLanguageContext and WolframAlphaContext in a single tool
• WolframAlphaContext - Semantic search of Wolfram|Alpha results (requires LLMKit)
• WolframLanguageContext - Semantic search of Wolfram Language documentation and other resources

Note: Since WolframContext combines the other two, a server should only include one of these three tools. Without LLMKit, WolframContext is effectively the same as WolframLanguageContext since the Wolfram|Alpha semantic search functionality is disabled.

Documentation search includes the Function Repository, Data Repository, Neural Net Repository, Paclet Repository, and more.

While only WolframAlphaContext requires an LLMKit subscription, having LLMKit greatly improves search results for all context tools by enabling reranking and filtering.

Code Execution Tools

• WolframLanguageEvaluator - Execute Wolfram Language code with time constraints
• WolframAlpha - Natural language queries to Wolfram|Alpha
• SymbolDefinition - Retrieve symbol definitions in readable markdown format

Notebook Tools

• ReadNotebook - Read Wolfram notebooks (.nb) as markdown text
• WriteNotebook - Convert markdown to Wolfram notebooks

Testing Tools

• TestReport - Run Wolfram Language test files (.wlt) and return reports

Code Analysis Tools

• CodeInspector - Inspect Wolfram Language code for issues and return formatted reports

Documentation Tools (Paclet Development)

• CreateSymbolDoc - Create new symbol documentation pages
• EditSymbolDoc - Edit existing symbol documentation pages
• EditSymbolDocExamples - Edit example sections of documentation

Paclet Release Tools (Paclet Development)

• CheckPaclet - Check a paclet for issues before building or submission
• BuildPaclet - Build a .paclet archive for distribution
• SubmitPaclet - Submit a paclet to the Wolfram Language Paclet Repository

MCP Apps Tools

• NotebookViewer - Embed interactive Wolfram Cloud notebooks inline
• MCPAppsTest - Diagnostic tool for testing the MCP Apps pipeline

Note: MCP Apps tools require a client that supports the io.modelcontextprotocol/ui extension. When UI support is available, the WolframAlpha and WolframLanguageEvaluator tools are automatically enhanced with interactive notebook viewers. See docs/mcp-apps.md for details.

See docs/tools.md for detailed information about each tool.

Creating Custom Servers

Create custom MCP servers with your own tools using LLMConfiguration:

config = LLMConfiguration[<|
    "Tools" -> {LLMTool["PrimeFinder", {"n" -> "Integer"}, Prime[#n]&]}
|>];

server = CreateMCPServer["My MCP Server", config]
(* Out: MCPServerObject[...] *)

Install for use in Claude Desktop:

InstallMCPServer["ClaudeDesktop", server]
(* Out: Success["InstallMCPServer", <|...|>] *)

After restarting Claude Desktop, your custom tools will be available:

You can also mix predefined tools with custom tools:

CreateMCPServer["My MCP Server", <|
    "Tools" -> {
        "WolframLanguageEvaluator",  (* Predefined tool *)
        "WolframAlpha",               (* Predefined tool *)
        LLMTool["MyCustomTool", ...]  (* Custom tool *)
    }
|>];

API Reference

Core Functions

Function	Description
CreateMCPServer[name, config]	Create a custom MCP server
InstallMCPServer[client]	Install the default server for a client
InstallMCPServer[client, server]	Install a specific server for a client
UninstallMCPServer[client]	Remove all servers from a client
UninstallMCPServer[client, name]	Remove a specific server from a client
DetectedMCPClients[]	Subset of $SupportedMCPClients whose config file exists on this machine
DeployAgentTools[target]	Deploy tools to a client with tracked deployment management
DeployAgentTools[All]	Deploy tools to every client in $SupportedMCPClients (see docs/deploy-agent-tools.md)
ValidateAgentToolsPacletExtension[paclet]	Validate an "AgentTools" paclet extension
DeployedAgentTools[]	List all tracked deployments
CreatePreferencesContent[]	Build the preferences panel UI for managing deployed Wolfram toolsets (see docs/preferences-content.md)

Deployment Objects

Symbol	Description
AgentToolsDeployment[...]	Data structure representing a tracked tool deployment
DeployedAgentTools[]	List all deployments
DeployedAgentTools[client]	List deployments for a specific client

Server Objects

Symbol	Description
MCPServerObject[...]	Data structure representing an MCP server
MCPServerObjectQ[expr]	Test if an expression is a valid server object
MCPServerObjects[]	List all created server objects (includes paclet-backed servers)

Predefined Resources

Symbol	Description
$DefaultMCPServers	Association of predefined server configurations
$DefaultMCPTools	Association of available tool definitions
$DefaultMCPToolOptions	Association of default tool option values
$DefaultMCPPrompts	Association of available prompt definitions

Options

CreateMCPServer

Option	Default	Description
OverwriteTarget	False	Overwrite an existing server with the same name
IncludeDefinitions	True	Include function definitions in the serialized server
Initialization	None	Code to evaluate when the server starts

InstallMCPServer

Option	Default	Description
"EnableMCPApps"	True	Enable or disable MCP Apps UI resources
"DevelopmentMode"	False	Use local source files instead of installed paclet
"MCPServerName"	Automatic	Override the config file key for the server entry
"ToolOptions"	<||>	Customize built-in tool behavior (see docs/tools.md)
"VerifyLLMKit"	True	Check LLMKit subscription requirements

Development

See the developer documentation for information on:

• Getting started with development
• Writing and running tests
• Building the paclet
• Adding new tools
• Agent skills for distributing tools to AI coding agents
• Paclet extensions for third-party tool/prompt/server contributions
• Error handling

For AI agents working on this codebase, see AGENTS.md.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

Richard Hennigan (Wolfram Research)