Stakpak is an open-source autonomous DevOps agent written in Rust that lives on the user's machines 24/7, keeps applications running, and only pings a human when intervention is required. Released under Apache-2.0, the project has 1,500+ GitHub stars, 147 forks, and active community traction as a security-hardened alternative to handing production credentials to general-purpose AI agents. Stakpak is positioned as a self-hosted PaaS replacement: all the upside of an opinionated deployment platform without the lock-in. ## Why Stakpak Matters Most AI coding agents today can produce DevOps code but cannot be trusted to operate infrastructure. A single hallucinated rm command, an overzealous terraform destroy, or an accidentally exposed secret can take down production. Stakpak addresses this by combining a long-running autonomous runtime with a security architecture explicitly designed for DevOps work: the LLM never sees raw credentials, destructive operations are blocked at the network layer, and a curated library of DevOps playbooks shapes the agent's default behavior. The result is an agent that can generate infrastructure code, debug Kubernetes, configure CI/CD, and automate deployments without being given the keys to production. ## Autopilot 24/7 Runtime Stakpak's Autopilot mode runs the agent as a persistent background service that monitors apps, executes scheduled jobs, and reacts to events on a continuous basis. Lifecycle is managed with simple commands - stakpak up to start the service, stakpak down to stop it, and stakpak autopilot doctor to check deployment readiness. Schedules are configured via cron expressions and can be tied to named profiles, and channels like Slack route notifications and approvals to the right humans. ## Security Hardened Architecture Stakpak ships with mutual TLS for end-to-end encrypted MCP, dynamic secret substitution that lets the agent read, write, and compare secrets without ever seeing their actual values, secure password generation with configurable complexity, and a privacy mode that redacts IP addresses, AWS account IDs, and other sensitive identifiers. The Warden Guardrails system enforces network-level policies that block destructive operations before they can reach production resources, providing a defense-in-depth layer that operates independently of agent reasoning. ## DevOps Capabilities The agent handles a broad surface of DevOps work out of the box: asynchronous task management for background processes like port forwarding and servers, real-time progress streaming for long-running Docker builds and deployments, infrastructure code indexing with semantic search across Terraform, Kubernetes manifests, Dockerfiles, and GitHub Actions, a documentation research agent for technical references and cloud provider docs, and bulk message approval for efficient multi-step workflows. File modifications are automatically backed up so any change can be reversed. ## Adaptive Intelligence with Rule Books Stakpak introduces Rule Books, a mechanism for customizing agent behavior with internal standard operating procedures, deployment playbooks, and organizational policies. The agent learns from interactions, remembers incidents, resources, and environment details, and adapts to specific team workflows over time. Specialized subagents can be enabled for code exploration and sandboxed analysis, each with different tool access levels appropriate to their task. ## Provider Flexibility Stakpak supports a Stakpak API key, Anthropic and OpenAI keys, or any custom OpenAI-compatible endpoint. The Docker image bundles the most common DevOps CLI tools - docker, kubectl, aws cli, gcloud, azure cli, and others - so the agent can act on cloud infrastructure without additional setup. Installation is available via Homebrew, binary releases, and a one-line curl installer on Linux, macOS, and Windows. ## Limitations Stakpak is focused on DevOps and infrastructure automation, so teams looking for a general-purpose coding agent will want to pair it with a separate IDE assistant. Autopilot requires Docker and at least 2GB of RAM on the host, and Linux user services may need linger configured to survive logout. As a relatively young project at 1,500+ stars, the rule book library and channel integrations are still expanding, and some advanced features such as subagents are gated behind feature flags.