MathWorks' Official Toolkit Connects Your Agent to MATLAB

Agents hallucinate MATLAB toolbox functions constantly — inventing signatures for functions that never existed, or missing the modern replacement MathWorks shipped last release. matlab/matlab-agentic-toolkit, the official MathWorks repository, attacks that at the root: it connects your AI agent to a live MATLAB session through the MATLAB MCP Server and equips it with curated skills for MATLAB workflows, conventions, and best practices.

Why This Skill Matters

This is the vendor's own packaging, so the knowledge comes from the same people who ship the toolboxes. The toolkit does two things: it auto-installs the MATLAB MCP Server, giving the agent tools to run MATLAB code and return command-window output, run tests via runtests with structured results, run static analysis with the Code Analyzer, and list the installed version and toolboxes (evaluate_matlab_code, run_matlab_file, run_matlab_test_file, check_matlab_code, detect_matlab_toolboxes). Two MCP resources add reference material: matlab_coding_guidelines and plain_text_live_code_guidelines.

On top of that sit the skills: 24 skill groups — 6 MATLAB groups (Core, App Building, Data Import and Analysis, External Language Interfaces, Programming, and Software Development) plus 18 toolbox groups spanning Aerospace, Control Systems, Image Processing and Computer Vision, Signal Processing, Radar, and Wireless Communications, among others. You do not need all 24; the README's guidance is to install only the groups relevant to your work, because agents trigger skills more reliably when fewer are loaded.

Installation

MATLAB R2021a or later is required. If you use Claude Code, GitHub Copilot CLI, or OpenAI Codex, the Agentic Toolkit Installer sets up the MATLAB and Simulink toolkits plus the MCP Server from one operating-system command line. From inside MATLAB instead:

1. Download the installer add-on (agenticToolkitInstaller.mltbx, linked in the README).
2. Open the downloaded file with MATLAB to install it.
3. Run:

setupAgenticToolkit("install")

4. Pick only the skill groups relevant to your work — re-run the installer later to add more.

By default your agent starts a new MATLAB session per call. To attach it to the session you already have open, run shareMATLABSession() in the MATLAB command window; if several sessions are running, the agent connects to the one where you ran it most recently.

Verify the wiring by asking your agent:

What version of MATLAB is running? List the installed toolboxes.

A real session answers with your actual version and toolbox list — the concrete answer that replaces the hallucinated one.

Real Workflow: Run and Test Code Through the Agent

With the toolkit connected, delegation becomes execution. Ask the agent to write a function that filters a timetable, then to run it — the call goes through evaluate_matlab_code and the command-window output comes back into the chat. Ask it to validate the work: run_matlab_test_file executes tests via runtests and returns structured results, and check_matlab_code runs the Code Analyzer before you ever commit. For guaranteed skill loading, trigger one by name — /matlab-write-tests in Claude Code.

Tips

• Updating is symmetric: download the latest installer add-on and run setupAgenticToolkit("update") — it refreshes skills, configuration, and the MCP server binary for both the MATLAB and Simulink toolkits.
• Working with Simulink? The companion Simulink Agentic Toolkit is offered at install time, and the same update command covers both.
• The README's security note asks you to review and validate every tool call and keep a human in the loop for important actions.
• The MATLAB MCP Server collects anonymized usage data by default; the documentation points to an opt-out if your environment forbids it.

When Not to Use This

You need a MATLAB license — the toolkit orchestrates MATLAB, it does not replace it, and MCP server use must comply with the MathWorks license agreement (no sharing a server across users). If your agent is not on the supported list (Claude Code, GitHub Copilot, OpenAI Codex, Gemini CLI, Amp), expect manual configuration of the MCP server and skills per your agent's own docs. And pull requests are not enabled on this repository — it is a distribution channel, so issues and support requests go through GitHub issues or MathWorks support.

See the leaderboard for more skills.