Blog
Introducing the Smart Classifier
What Is the Smart Classifier?
The Smart Classifier is actower's built-in system for categorizing the questions and tool requests that AI agents generate. Instead of reviewing every single prompt, you can let the classifier sort them for you.
How It Works
When an agent asks a question or requests permission to use a tool, actower's classifier analyzes the request and assigns one of 20+ types. The most common:
| Type | Description |
|---|---|
`BASH_SAFE` | Read-only shell commands (ls, cat, grep) |
`BASH_RISKY` | Write or destructive commands (rm, chmod, mv) |
`READ_FILE` | File read operations |
`WRITE_FILE` | File write or edit operations |
`MCP_TOOL` | MCP server tool calls |
Automatic Auto-Approval
Once a request is classified, actower can respond for you. Its built-in policy approves the safe, read-only operations instantly — types like BASH_SAFE and READ_FILE — while risky ones like BASH_RISKY and WRITE_FILE are held for your approval.
The safe/risky split is built in: actower classifies each request and already knows which types are safe to approve automatically. Your agents move at full speed on the safe stuff, and you only stop for what matters.
2,900+ Real-World Patterns
The classifier is regression-tested against over 2,900 real-world patterns captured from actual agent sessions. It handles edge cases like:
- Piped commands (
cat file | grep pattern) - Compound commands (
mkdir -p dir && cd dir) - Subshells and redirections
- Common aliases and wrappers
Try It Out
Enable classification with a single flag:
actower monitor --all --classifyYou'll see each question annotated with its type in both the TUI and Web UI.