Dev Tools · 2h ago
Open-source MCP server lets AI run Linux commands with a human consent gate
A developer built devops-mcp, an open-source MCP server that grants AI assistants like Claude Desktop read-only access by default and requires a secret token for any state-changing command. It offers three trust levels (safe, provision, full) with automatic expiry and blocks irrecoverable operations unless a backup is confirmed. The tool logs every command and resists prompt injection by tagging server output as data, not instructions.
Meridian48 take
This is a practical, well-thought-out safety layer for AI-assisted DevOps, but its real-world adoption will depend on how easily it integrates into existing workflows and whether teams trust the threat model enough to use it on production systems.
Read the full reporting
I gave Claude SSH access to my server — here's the consent gate that makes it safe →
DEV Community
mcp-serverai-safety