Salvager Launches Filesystem-Level Undo Tool for AI Coding Agents

Summary

Salvager is a new open-source safety tool that provides automatic file revision tracking for AI coding agents. The tool passively captures every file change—whether from human edits or agent modifications—into a hidden .salvager/ directory, allowing developers to recover previous versions at any point. With zero configuration required and a single static Go binary, Salvager integrates seamlessly into existing workflows while offering recovery options via CLI, MCP (Model Context Protocol), or manual file inspection.

The tool addresses a critical gap in AI agent safety: mistakes happen fast and silently. An agent can rewrite a file faster than a developer can review the diff, and by the time issues are discovered, the original version may be unrecoverable through standard git or filesystem tools. Salvager operates at the filesystem level, independent of version control or editor state, ensuring protection even when agents work outside conventional workflows.

Key features include automatic per-file revision capture, zero-config setup, read-only MCP endpoints that allow agents to self-recover their mistakes, and guaranteed reversibility—every restore operation itself becomes a new revision. The tool is designed to be trustworthy: it captures everything, requires no accounts or cloud infrastructure, and critically, exposes no delete functionality over MCP, preventing agents from erasing their own safety record.

• Reversible recovery design ensures nothing is a dead end; every action is undoable and auditable
• Open-source and fully readable; history stored as plain files in .salvager/ directory

Editorial Opinion

Salvager addresses a genuinely novel problem in the AI coding agent space: the ability for agents to make irreversible mistakes faster than humans can catch them. By placing recovery at the filesystem level—independent of git, editors, or the agent itself—it sidesteps the usual weak points in safety tools. The MCP integration that lets agents self-recover is particularly clever, treating the agent as an intelligent participant in recovery rather than just a liability. For teams deploying AI agents in production, this kind of foundational safety infrastructure may become as essential as version control itself.