Show HN: DashClaw – policy and approval layer for unattended coding agents (github.com)

DashClaw, a newly announced policy and approval layer for unattended coding agents, is designed to enhance governance and safety in AI-driven software development. It acts as a fail-closed interceptor, halting potentially destructive or costly actions before execution by routing them to an approval inbox. This ensures that developers maintain control over their agents, even when they’re not actively monitoring the coding process. The system records every decision in a signed ledger, providing an auditable trail that includes specific commands and the rationale behind approvals or denials. Significantly, DashClaw enables solo developers or small teams to run long, unattended coding sessions without forfeiting oversight, thus protecting their infrastructure from accidental misuse. It features a user-friendly interface for remote approvals and customizable policies, calibrating how often interventions occur while allowing agents to operate with defined levels of autonomy. The introduction of prompt-injection scanning and decision scoring marks a serious advancement in embedding safety mechanisms directly within AI workflows. This development not only promises to mitigate risks associated with unattended automation but also sets a precedent for responsible AI usage in software development.