GuardBench — Local guardrail regression harness for coding agents

Turn product constraints into regression tests so Claude Code can’t overstep

Problem

Critical product constraints often live only in an engineer’s or PM’s head. When coding agents run unsupervised, they easily cross red lines—touching files they shouldn’t or calling APIs they shouldn’t. A developer built a local guardrail layer specifically for this, showing it’s a widespread anxiety when using coding agents.

Solution

Let users describe allowed/forbidden actions in natural language or a simple DSL, generating a local set of enforceable guardrails. Rules load automatically before each agent run, block violations during execution, and log everything. A regression report shows which rules fired and which constraints were bypassed.

Users

Engineers and tech leads using Claude Code, Codex, or similar agents in teams, especially projects with compliance or code-ownership requirements.

MVP

CLI to define guardrails (file paths, commands, network access); inject local check hooks before an agent run; block violations and write logs; generate a one-page regression report.

Revenue

Open-source CLI; team plan subscription per seat for shared policies, audit logs, and Slack alerts.

Why now

Coding agents are shifting from autocomplete to executors that touch large codebases, yet explainable constraint tooling is still primitive—an easy wedge for indie developers.