The latest release of mcpware/intentprobe is architecture-v0.1.2. This Python research repository covers, for AI agent classification and guardrails, an architecture in which multiple task-specific probes read together the hidden states of a local language model that is fixed and run once. The probes each take on different detection tasks without repeating the model execution. The repository states that it is licensed under Apache-2.0 and provides public models and code along with reproducible benchmarks.
In this architecture, detection results, policy decisions, and actual execution authority are separated. A probe therefore reports what it detected in the input, while the policy decision or the execution based on that result is handled by a separate layer. Each small probe learns and is versioned independently. The materials include a description of a probe that classifies whether the user approved something, and not only a public scanner and a single probe but also an execution path that links several probes, experiment logs, reproduction materials, and tests. This is closer to a research tool that examines the approach of composing detectors by splitting them up, than to a finished single classifier.
That said, the performance comparisons in the README are research measurements and cannot be read as general evaluation rankings or product-level performance. The reproduction materials and tests provide clues for reviewing the structure, but performance comparisons alone do not confirm effectiveness in a real agent environment or suitability for deployment. It is therefore more reasonable to treat the grounds for this repository as limited to the scope of the probe-sharing architecture and the experimental materials.