The MATLAB toolkit's installer is downloaded from the Simulink toolkit's releases

Step one of the install points at another product's release assets

The badges at the top of the page link to this repository's own latest release, twice. The first installation step then tells you to download an installer add-on from a URL under a different repository, the Simulink agentic toolkit, using that project's releases endpoint and a file named for the installer. So the MATLAB toolkit's setup depends on the Simulink repository continuing to publish that asset under that name, and a reader who assumes the two products are independent will not find the file by looking at the repository they just cloned. The file is an add-on archive that you open in MATLAB to install as an add-on, so the download is not optional in the way a script would be.

Three releases with three different naming schemes

The release history is short and inconsistent. The newest is a product-prefixed calendar version with a letter suffix, dated late September 2026, and the one before it is the same scheme with the letter a, dated early September. The third is a bare date, with no product prefix and no letter, dated the same day it was published. So the scheme appears to have changed and then changed back, or the third tag predates a convention that was introduced afterwards. That matters more than it would in a small project, because the repository root carries a citation file, which is exactly the kind of artefact that assumes a stable version string. The last push on the repository is dated 2026-09-30, six days after the newest release.

The agent gets a live session and five tools, one of which runs code

The tool table is the capability list and it is small on purpose. One tool runs MATLAB code and returns the command window output, one runs a program file, one runs a test suite and returns structured results, one does static analysis with the Code Analyzer, and one reports the installed version and the toolboxes present. Alongside them sit two resources, one holding MATLAB coding standards and one holding Live Script format rules, which the agent reads when it needs them rather than carrying them in context all the time. The first tool is the one to think about: an agent that can evaluate code in your session can change your workspace, and the guard rails here are the session boundary rather than a permission prompt.

With several sessions open, the agent attaches to the most recent one

By default the agent creates a new MATLAB session when you call it. To change that, you run one command in the MATLAB command window, and the page explains exactly what happens when more than one session is open: the agent connects to the session where you most recently ran that command. The two commands you actually type are short. Install first:

setupAgenticToolkit("install")

And the session attach is one line:

shareMATLABSession()

There is no list to choose from and no session identifier to pass, so a background session you started later becomes the one your agent drives. The command can also be added to the MATLAB startup script, which means it will run in every session you open from then on. This is the single most important line in the installation section for anyone working with a long-lived interactive session, and it is two lines long.

The vendor says fewer skills make skill selection more reliable

The instructions are unusually candid about a failure mode of this approach. Step four of the installation tells you to install only the skill groups relevant to your work, and gives a reason: it helps your agent reliably trigger the right skills. The skills section repeats it, saying agents are more reliable at triggering skills when fewer are loaded. You can also bypass selection entirely by naming a skill, with a slash-command form given as the example, so there is a manual override when the agent picks wrong. Seven MATLAB-level groups are listed, from core code and app building through data analysis, environment migration, external language interfaces, programming practices and software development, plus a second table mapping toolbox groups to supported products.

The floor is a 2021 release, and one skill group exists because of that

The stated requirement is MATLAB R2021a or later, alongside an AI coding agent that supports MCP servers and skills. Five agents are named as supported and are configured automatically: Claude Code, GitHub Copilot, OpenAI Codex, Gemini CLI and Amp. A floor that far back is a deliberate choice about reach, and the toolkit carries the cost of it in its own catalogue: one group is specifically for diffing settings between releases and migrating startup scripts to correct setting paths, which is the work you would have to do to move across exactly that range. Another group covers upgrading MEX files to the interleaved complex API, which is a similar migration-shaped task.

Verification is done by asking the agent, and the licence is a file only

The verify step is to ask your agent which version of MATLAB is running and to list the installed toolboxes. There is no command in the page for you to run yourself, which means the toolkit's own smoke test exercises the model context path rather than confirming the add-on independently, so a broken server and a broken agent look the same from the user's side. The longer document linked from the top covers the cases that page skips: installing from local files, installing in an offline environment, configuration options, platform-specific notes, verification, troubleshooting, and manual setup without the installer. Licensing is the other loose end, since the repository's licence field is unresolved while a licence file sits at the root.

Editorial conclusion

The design here is the part worth copying. Rather than trying to teach a model MATLAB from documentation, the vendor ships curated skills per product, keeps the surface deliberately small with five tools and two reference resources, and tells you to install fewer skill groups because skill selection degrades as the list grows. The MCP surface is also honest about what it is: an agent that can evaluate code in a live session, run tests, and call the static analyser. Two things to check before you rely on it. Step one of the installation downloads the installer add-on from the Simulink toolkit's release assets rather than its own, so the two products are coupled through a URL. And the session model defaults to a new MATLAB session, with an explicit command to attach to the one you have, and with several open the most recent one wins. Read the session behaviour before pointing an agent at a long-running session.

Frequently asked questions

Official sources

1. Issues
2. matlab/matlab-agentic-toolkit on GitHub
3. Project website
4. README
5. Releases