PromptGuard: AI Output Safety Middleware

A plug-and-play safety layer for AI coding agents, built for indie developers.

Problem

AI coding assistants like Claude Code can run dangerous commands (e.g., `rm -RF`) or load untrusted dependencies on a developer's machine. Existing solutions are either full sandboxes (e.g., SideKernel, requiring microVMs) or simple blocklists that lack fine-grained, auditable interception in real workflows. Users also worry about sensitive data (keys, internal paths) leaking through AI outputs without notice.

Solution

A lightweight local middleware (CLI + menubar/system service) that inserts a rule engine between AI coding tools and the OS: intercepts dangerous commands via regex and AST patterns, scans tool outputs for secrets/PII, enforces allowlists for file and network access, and logs all interceptions for auditing. Rules are user-configurable via TOML/YAML, no VM required.

Users

Indie developers and small team leads using Claude Code, Cursor, Copilot Workspace, who care about local security without sacrificing speed.

MVP

CLI wrapper around AI coding tools that parses stdout/stderr and intended commands; 20 built-in dangerous patterns (rm/mv/chmod etc.); redact and alert on suspected secrets in outputs; write all events to local SQLite for review; simple web UI for logs and rule editing.

Revenue

Open-core + paid rule subscription (weekly updated dangerous patterns & sensitive data regex library); or team seat subscription for advanced audit and compliance export.

Why now

AI coding agents are shifting from IDE-embedded to autonomous CLI execution, creating a huge local security gap; existing sandboxes are too heavy, making middleware a lightweight entry point for indie developers.