Prerequisites¶
Before running any cube agentic-ai command, the following must be in place on your machine.
cube bundle manager¶
The installer is a Cube bundle. The cube CLI must be installed and available on PATH.
To verify:
cube --version
If cube is not found, complete the bootstrap first:
https://dev.st.com/stm32cube-agentic-ai/stm32cube-agentic-ai-start/index.html
Installer bundle¶
The installer bundle must also be installed. If bootstrap did not install it, run:
cube bundle --yes install cube-agentic-ai-installer
cube agentic-ai --version
The installer uses its Cube-managed Node.js runtime. A separate system Node.js installation is not required. Before deploying contexts, create or choose an existing project directory. The folder does not need to contain generated firmware or a Cube project yet.
Operating system¶
OS |
Tested versions |
Architecture |
|---|---|---|
Windows® |
Windows® 10 and 11, 64-bit |
x86_64 |
Linux® |
Ubuntu® 22.04, 24.04, 26.04 — Fedora® 43, 44 |
x86_64 |
macOS® |
macOS® 15 and 26 |
Intel® (x86_64) and Apple® Silicon (aarch64) |
Warning
Ubuntu® 20.04 is not supported due to an incompatible glibc version.
Disk space¶
At least 500 MB free disk space in the bundle installation directory.
The exact location depends on your OS and the bootstrap configuration. Run cube --get-current-value CUBE_BUNDLE_PATH to find it on your machine.
Network access¶
Network access is required for:
Downloading bundles from the Cube registry during
machine-installandmachine-bootstrap.Downloading additional Python packages, if declared by a selected service, during
setupanddoctor --fix, from the configured pip package index (proxy, certificate authority, mirror).Connecting to remote MCP server endpoints declared by installed contexts at runtime.
Offline operation is possible for show-catalog, list-supported-targets, project-lock-show,
and project-deploy-installed when the required local bundles and their dependencies are already
available. Remote MCP services still need network access when used by the agent.
Each context can have additional prerequisites, such as editor extensions or tools not declared as
bundle dependencies. Installing a context does not install these automatically. cube agentic-ai setup
and cube agentic-ai doctor check the ones that the context declares and report what to do: the
remediation of a prerequisite, or the code --install-extension command of an extension.
GitHub Copilot in VS Code requires the GitHub Copilot extension, and Claude
Code in VS Code requires Anthropic’s Claude Code extension. Install the
appropriate agent extension yourself. The installer does not install or manage
either one. The optional --install-vscode-extensions flag applies only to
extensions declared by toolkit contexts. See the GitHub Copilot extension
Marketplace page
and the Claude Code for VS Code guide.
The eSW HAL/LL guidance service requires Python 3.10 or later. The connector does not install
Python. setup checks the interpreter and can install additional packages declared by a selected service. Use
setup --python auto (default), --python venv, or --python "<absolute Python path>" when setup
creates the project. The choice of a deployed project is kept. The selection
is recorded per project and ESW bundle version under .cube-ai/esw-python/. Projects selecting
venv for the same ESW version share that bundle version’s environment. See Python of the eSW
context.
Build and debug contexts require VS Code and declare their extensions as dependencies. By default,
setup reports the commands for the selected VS Code profile. The installer cannot inspect an active
profile from the terminal. setup --install-vscode-extensions is optional and requires the code
command on PATH. Otherwise, install from the Extensions view.
PATH¶
The directory where cube is installed must be on your shell PATH.
The bootstrap script adds this automatically.
On Windows, open a new PowerShell window after installation for the change to take effect.