Your infrastructure's night shift, now with guardrails

What it does Stakpak is a Rust-based autonomous agent that runs continuously on your own hardware to manage deployments, generate infrastructure code, debug Kubernetes, and keep applications running. It operates through a terminal UI or a background autopilot mode that supports cron schedules and Slack channels. You can point it at commercial LLMs or run it fully offline through local endpoints like Ollama.

The interesting bit The project treats LLM security as a first-class concern rather than an afterthought: it substitutes secrets dynamically so the model never sees plaintext credentials, enforces network-level guardrails to block destructive operations, and wraps tool access in mutual TLS when acting as an MCP server or proxy. That combination of autonomous shell access and enforced ignorance is unusual in the current crop of coding agents.

Key highlights

• 24/7 autopilot mode with cron schedules and channel integrations for unsupervised operations
• Dynamic secret substitution and privacy-mode redaction for IPs and AWS account IDs
• Local semantic search across Terraform, Kubernetes, Dockerfile, and GitHub Actions code
• Reversible file operations with automatic backups, plus bulk approval for multiple tool calls
• Can act as a secure MCP proxy or server with local-only, remote-AI, or combined tool modes

Caveats

• Autopilot mode requires Docker, recommends 2 GB+ RAM, and may need extra Linux configuration like swap and linger to survive logout
• Background updates require manually restarting long-running processes to pick up the new binary
• The Stakpak API key flow warns that “brave users may encounter issues with automatic redirects to localhost ports”

Verdict Worth a look if you want PaaS-like automation on your own infrastructure without cloud lock-in. Skip it if you prefer point-and-click UIs or aren’t comfortable giving an agent shell access, however sandboxed.