Getting Started with the MATLAB® Agentic Toolkit

May 6, 2026 · View on GitHub

This guide walks you through setup of the MATLAB Agentic Toolkit. Once complete, the toolkit will be configured globally, for you to use on any project.

For a project overview, tool/skill reference, and documentation links, see the README.

Automated setup has been verified with basic workflows on each platform. The toolkit is under active development — please report issues if you encounter problems.

Prerequisites

Before you begin, make sure you have:

MATLAB R2020b or later installed
An AI coding agent that supports MCP — see Supported Platforms
Git™ (to clone the toolkit)

Choose Your Path

Path	When to use	What you get
Full Setup	First time; want everything automated	MCP server binary + skills + global configuration
Adding Skills Only	Already have the MCP server configured	Skills/prompts only; no MCP changes

Full Setup

Full setup clones the toolkit repository, then uses your agent to automate the entire configuration. This is the recommended path for first-time users.

What Setup Does

Installs the MCP server — downloads the MATLAB MCP Core Server binary to ~/.matlab/agentic-toolkits/bin/
Configures your agent — connects the MCP server to your agent via global config, defaulting to the most recent MATLAB found (you can change this during setup)
Registers skills — adds MATLAB skills via the platform's native plugin system or global skill links
Verifies — confirms the MCP server can reach MATLAB and reports an environment summary

Setup is re-runnable. Run it again to upgrade the MCP server, switch MATLAB versions, or fix a broken configuration.

Clone the toolkit to a permanent location outside your project directories (e.g., ~/tools/ or ~/repos/). Most platforms reference this clone via symbolic links, so the toolkit needs to stay in place after setup.

git clone https://github.com/matlab/matlab-agentic-toolkit.git
cd matlab-agentic-toolkit

Then launch your agent:

Platform	Launch command
Claude Code	`claude`
OpenAI Codex	`codex`
Gemini CLI	`gemini`
GitHub Copilot	Open folder in VS Code, then start Copilot chat
Sourcegraph Amp	`amp`

Step 2: Run Setup

Ask your agent:

Set up the MATLAB Agentic Toolkit

The setup skill walks you through detection, planning, and execution. It presents all decisions (which MATLAB to use, display mode, etc.) before making any changes.

After setup completes, start a new session in any project directory — MATLAB tools and skills are available everywhere.

Tip: After the first setup, you can re-run setup at any time to upgrade the MCP server, switch MATLAB versions, or fix a broken configuration. In Claude Code, use the /matlab-setup slash command. In other agents, ask "Set up the MATLAB Agentic Toolkit" from the toolkit directory.

How Configuration Works per Platform

Setup writes two things: an MCP server configuration (so your agent can talk to MATLAB) and skill registrations (so your agent has MATLAB expertise). The details vary by platform. For more on how skills update, see Updating.

Platform	MCP Configuration	Skills Delivery	How To Update Toolkit
Claude Code	`~/.claude/settings.json`	Plugin cache	Re-run setup or reinstall plugin
GitHub Copilot	VS Code user-profile `mcp.json`	`~/.agents/skills/` symlinks	`git pull` in toolkit repo, re-run setup
OpenAI Codex	`~/.codex/config.toml`	`~/.agents/skills/` symlinks	`git pull` in toolkit repo, re-run setup
Gemini CLI	`~/.gemini/settings.json`	`~/.agents/skills/` symlinks	`git pull` in toolkit repo, re-run setup
Sourcegraph Amp	`~/.config/amp/settings.json`	`amp.skills.path` direct ref	`git pull` in toolkit repo, re-run setup

How skill symlinks work: Most platforms discover skills from ~/.agents/skills/. Setup creates symbolic links from that directory to the individual skill directories in your toolkit clone. When you run git pull, the linked skills update automatically. If new skills are added to the toolkit, re-run setup to create the additional symlinks.

Where we're headed: As agent platforms mature their marketplace and plugin systems, the goal is to move toward a simple marketplace-based install for all platforms — add the MATLAB Agentic Toolkit marketplace, install the MATLAB Core plugin, and the plugin handles everything including MCP server installation and configuration. The current clone-and-setup workflow is expected to be a temporary work-around to help provide a good out-of-the-box experience in the meantime.

Platform-Specific Notes

Claude Code — Setup registers the toolkit via the plugin marketplace and uses claude mcp add-json -s user to register the MCP server globally. Skills are cached by the plugin system. To re-run setup, use /matlab-setup or ask Claude to set up the toolkit.

GitHub Copilot — Setup writes global MCP config to the VS Code user-profile mcp.json (~/Library/Application Support/Code/User/mcp.json on macOS, ~/.config/Code/User/mcp.json on Linux, %APPDATA%\Code\User\mcp.json on Windows) and creates skill symlinks in ~/.agents/skills/. Reload VS Code after setup completes (Cmd/Ctrl + Shift + P, then "Developer: Reload Window").

OpenAI Codex — Setup uses codex mcp add if available, otherwise writes ~/.codex/config.toml directly. Skills are installed as global symlinks in ~/.agents/skills/. After setup, you may want to tune two settings in the [mcp_servers.matlab] section of ~/.codex/config.toml:

tool_timeout_sec = 600 — increases the tool timeout from the default (which is too short for many MATLAB operations like test suites and simulations). Increase further for very long-running tasks.
env_vars = ['WINDIR'] — Windows only. Required for Simulink to work, since Codex strips environment variables from MCP server subprocesses by default.

Gemini CLI — Setup writes global config to ~/.gemini/settings.json and creates skill symlinks in ~/.agents/skills/. Start a new Gemini session after setup. Alternatively, you can install the toolkit as a Gemini extension (see Installing as a Gemini Extension below).

Sourcegraph Amp — Setup writes to ~/.config/amp/settings.json using the amp. prefix for all keys. Skills load directly from the toolkit via amp.skills.path (no symlinks needed). If you have amp.mcpPermissions rules that block MCP servers, setup will detect this and ask before making changes.

Installing as a Gemini Extension

You can install the toolkit as a Gemini CLI extension. The extension does not include MCP server configuration (MCP config is system-specific — binary path, MATLAB root, display mode), so you configure MCP separately.

Install and configure the MCP server following the steps in Installing the MCP Server Manually, then add it to ~/.gemini/settings.json:

{
  "mcpServers": {
    "matlab": {
      "command": "/absolute/path/to/matlab-mcp-core-server",
      "args": ["--matlab-root", "/absolute/path/to/MATLAB/R20XXx", "--matlab-display-mode", "desktop"]
    }
  }
}

Install the extension:

gemini extensions install https://github.com/matlab/matlab-agentic-toolkit

Create skill symlinks so Gemini discovers the toolkit skills:

mkdir -p ~/.agents/skills
for skill in ~/.gemini/extensions/matlab-agentic-toolkit/skills-catalog/matlab-core/*/; do
  ln -s "$skill" ~/.agents/skills/$(basename "$skill")
done

Start a new Gemini session and verify: ask "What version of MATLAB is running?"

Tip: For a fully automated setup that handles all of the above, use the Full Setup workflow instead.

Adding Skills Only

If you already have the MATLAB MCP Core Server installed and configured, you only need skills.

Claude Code (no clone required)

Add skills directly via the plugin marketplace:

claude plugin marketplace add "https://github.com/matlab/matlab-agentic-toolkit"
claude plugin install matlab-core@matlab-agentic-toolkit

Choose your preferred scope (per-project, per-user, or global) when prompted. Your existing MCP configuration is not modified.

To also get the setup skill (for managing the MCP server later), additionally run: claude plugin install toolkit@matlab-agentic-toolkit

Other Platforms

Point your agent's skill or prompt directory at skills-catalog/matlab-core/. Each skill is a self-contained SKILL.md with a manifest.yaml.

For platforms that discover skills from ~/.agents/skills/, create symlinks:

mkdir -p ~/.agents/skills
for skill in /path/to/matlab-agentic-toolkit/skills-catalog/matlab-core/*/; do
  ln -s "$skill" ~/.agents/skills/$(basename "$skill")
done

Replace /path/to/matlab-agentic-toolkit with the actual path to your toolkit clone.

Installing the MCP Server Manually

If you prefer to install the MCP server yourself rather than using the automated setup, follow these steps.

1. Download the binary

Go to the MATLAB MCP Core Server releases page and download the binary for your platform:

OS	Architecture	Asset Name
macOS	Apple silicon (arm64)	`matlab-mcp-core-server-maca64`
macOS	Intel (x86_64)	`matlab-mcp-core-server-maci64`
Linux	x86_64	`matlab-mcp-core-server-glnxa64`
Windows	x86_64	`matlab-mcp-core-server-win64.exe`

2. Install the binary

Place the downloaded binary in ~/.matlab/agentic-toolkits/bin/, rename it to matlab-mcp-core-server (or matlab-mcp-core-server.exe on Windows), and make it executable:

macOS:

mkdir -p ~/.matlab/agentic-toolkits/bin
mv ~/Downloads/matlab-mcp-core-server-maca64 ~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server
chmod +x ~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server
xattr -d com.apple.quarantine ~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server 2>/dev/null

Linux:

mkdir -p ~/.matlab/agentic-toolkits/bin
mv ~/Downloads/matlab-mcp-core-server-glnxa64 ~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server
chmod +x ~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server

Windows (PowerShell):

New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.matlab\agentic-toolkits\bin"
Move-Item ~\Downloads\matlab-mcp-core-server-win64.exe ~\.matlab\agentic-toolkits\bin\matlab-mcp-core-server.exe
Unblock-File -Path "$env:USERPROFILE\.matlab\agentic-toolkits\bin\matlab-mcp-core-server.exe"

3. Verify the binary runs

~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server --version

macOS note: If macOS blocks the binary (Gatekeeper), go to System Settings > Privacy & Security and click "Allow Anyway" next to the blocked binary.

4. Configure your agent

Point your agent's MCP configuration at the installed binary. See the How Configuration Works per Platform table for configuration file locations, or consult your agent's documentation.

Verification

Check that skills are loaded

If your agent shows loaded skills or plugins in its UI (e.g., Claude Code's /skills command), confirm the MATLAB Agentic Toolkit skills are listed. See the skills catalog for the full list of available skills.

Try it out

Ask your agent:

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

The agent calls detect_matlab_toolboxes via MCP and reports the MATLAB version and available toolboxes.

More examples

Write a function that computes the moving average of a signal, then generate unit tests for it.

Review the file myScript.m for code quality issues and suggest improvements.

Create a plain-text Live Script that demonstrates curve fitting with sample data.

Updating

The toolkit has two independently versioned components: the toolkit itself (skills, configurations, and documentation in this repository) and the MCP server binary (MATLAB MCP Core Server). Update each one separately.

Updating skills and configurations

Pull the latest changes from the repository:

cd /path/to/matlab-agentic-toolkit
git pull

What happens after git pull:

Platform	Effect
Copilot, Codex, Gemini CLI	Existing skills update immediately (symlinks point into the repo). If new skills were added, re-run setup to create the new symlinks.
Sourcegraph Amp	Skills update immediately (`amp.skills.path` reads the repo directly), including new skills.
Claude Code	Skills may be cached by the plugin system — re-run setup or reinstall the plugin to refresh.

Updating the MCP server

The MCP server binary is released independently from the toolkit. To update it:

Automated: Re-run setup (e.g., /matlab-setup in Claude Code, or ask any agent to set up the toolkit). Setup detects the installed version and downloads a newer release if available.
Manual: Download the latest binary from MATLAB MCP Core Server releases and replace the existing binary. See Installing the MCP Server Manually for platform-specific instructions.

Tip: Watch or star the MATLAB MCP Core Server repository on GitHub to get notified of new releases.

Per-Project Configuration

Automated setup configures the toolkit globally — MATLAB tools and skills are available in every session regardless of which project you open. Global configuration is the easiest way to get started since it just works whenever you create or open a project.

You can also configure the MCP server at the project level. This keeps your tools and skills scoped to the projects that need them, and it helps teams — when the config is committed to version control, anyone who clones the repo gets the MATLAB connection automatically (provided they have the MCP server binary installed; see Installing the MCP Server Manually).

Template files

The templates/ directory contains starter configurations for each platform. Copy the appropriate template into the root folder of your project, update the paths, and commit it to version control.

Platform	Template	Project location
GitHub Copilot	`templates/vscode-mcp.json`	`.vscode/mcp.json`
Sourcegraph Amp	`templates/amp-settings.json`	`.amp/settings.json`
OpenAI Codex	`templates/codex-mcp.json`	`.codex/config.json` in project root

Claude Code uses claude plugin install with scope selection (per-project, per-user, or global) rather than a project config file. See Adding Skills Only.

Example: GitHub Copilot

mkdir -p .vscode
cp /path/to/matlab-agentic-toolkit/templates/vscode-mcp.json .vscode/mcp.json

Then edit .vscode/mcp.json to replace the placeholder paths with your actual MCP server binary and MATLAB root paths.

Note: Per-project configs contain absolute paths to the MCP server binary and MATLAB root, which vary across machines. If your team uses different OS platforms or install locations, consider documenting the expected paths in your project README.

Troubleshooting

Problem	Likely Cause	Fix
Setup can't find MATLAB	Non-standard install location	Provide the path when prompted
MCP server download fails	Network/proxy/firewall	Download manually from GitHub releases, place in `~/.matlab/agentic-toolkits/bin/`, re-run setup
macOS blocks the MCP server binary	Gatekeeper quarantine	Setup handles this automatically. If still blocked (MDM), go to System Settings > Privacy & Security > Allow Anyway
Agent doesn't list MATLAB skills	Plugin not installed or skills not linked	Re-run setup; for Claude Code, try `claude plugin install matlab-core@matlab-agentic-toolkit`
MCP tools fail to connect	MCP server binary missing or wrong path in configuration	Re-run setup to regenerate configuration. Verify binary exists: `~/.matlab/agentic-toolkits/bin/matlab-mcp-core-server --version`
`evaluate_matlab_code` returns errors	Wrong `--matlab-root` path, license issue, or MATLAB startup failure	Verify MATLAB can start: `<matlab-root>/bin/matlab -nodesktop -r "disp('ok'),quit"`. Check license status. Re-run setup to correct the MATLAB root path
Codex tool calls time out	Default tool timeout too short for MATLAB	Add `tool_timeout_sec = 600` (or higher) to `[mcp_servers.matlab]` in `~/.codex/config.toml`
Simulink fails in Codex on Windows	Missing `WINDIR` environment variable	Add `env_vars = ['WINDIR']` to `[mcp_servers.matlab]` in `~/.codex/config.toml`

MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See mathworks.com/trademarks for a list of additional trademarks. Other product or brand names may be trademarks or registered trademarks of their respective holders.