Summary
When Claude (or any MCP-connected AI agent) attempts to use a MATLAB function that requires an Add-On support package, it currently receives a hard error and falls back to a less accurate alternative — rather than installing the missing package and retrying. This degrades output quality silently and breaks the "just works" expectation of agentic workflows.
Current Behavior
When an agent calls a function backed by an uninstalled support package, MATLAB returns an error like:
peopleDetector requires the Computer Vision Toolbox Model for RTMDet Object Detection
support package. To install this support package, use the Add-On Explorer.
Because the Add-On Explorer is a GUI action, the agent cannot act on this link. In the example below, Claude initially attempted to use peopleDetector() (RTMDet-based, higher accuracy), hit this error, and silently fell back to the legacy HOG-based vision.PeopleDetector:
● matlab - evaluate_matlab_code → Error: peopleDetector requires the Computer Vision
Toolbox Model for RTMDet Object Detection support package. To install this support
package, use the Add-On Explorer.
● The default peopleDetector needs a support package. Let me use the classic
HOG-based vision.PeopleDetector instead.
The agent never informed the user that it had downgraded its approach. The user asked for the best available method; they got a legacy fallback.
Desired Behavior
The MCP server should expose a mechanism for agents to:
- Detect that a function call failed due to a missing support package (parse the specific error pattern)
- Identify the required package name from the error message
- Install it programmatically — e.g. via
matlab.addons.install(...) or matlab.addons.supportpackage.install(...)
- Retry the original function call automatically
Optionally, the server could expose a dedicated tool (e.g. install_support_package) so agents can make the install decision explicitly and transparently before executing code.
Why This Matters
- Agentic workflows are non-interactive by design — GUI prompts are a dead end
- Silent capability degradation is worse than a clear error; users lose trust when they discover the agent quietly used a worse method
- Support package installation is already scriptable in MATLAB; the MCP layer just needs to surface it
Suggested API / Tool
{
"tool": "install_support_package",
"params": {
"package_id": "RTMDET_DETECTION"
}
}
Or as automatic middleware: intercept support-package errors before returning them to the agent, attempt a silent install, and retry once.
Environment
- MATLAB MCP Core Server
- Triggered via: Claude (claude-sonnet), agentic tool use
- Example function affected:
peopleDetector() (requires RTMDET_DETECTION)
Summary
When Claude (or any MCP-connected AI agent) attempts to use a MATLAB function that requires an Add-On support package, it currently receives a hard error and falls back to a less accurate alternative — rather than installing the missing package and retrying. This degrades output quality silently and breaks the "just works" expectation of agentic workflows.
Current Behavior
When an agent calls a function backed by an uninstalled support package, MATLAB returns an error like:
Because the Add-On Explorer is a GUI action, the agent cannot act on this link. In the example below, Claude initially attempted to use
peopleDetector()(RTMDet-based, higher accuracy), hit this error, and silently fell back to the legacy HOG-basedvision.PeopleDetector:The agent never informed the user that it had downgraded its approach. The user asked for the best available method; they got a legacy fallback.
Desired Behavior
The MCP server should expose a mechanism for agents to:
matlab.addons.install(...)ormatlab.addons.supportpackage.install(...)Optionally, the server could expose a dedicated tool (e.g.
install_support_package) so agents can make the install decision explicitly and transparently before executing code.Why This Matters
Suggested API / Tool
{ "tool": "install_support_package", "params": { "package_id": "RTMDET_DETECTION" } }Or as automatic middleware: intercept support-package errors before returning them to the agent, attempt a silent install, and retry once.
Environment
peopleDetector()(requiresRTMDET_DETECTION)