agent

Stakpak Agent is an open-source agent positioned as "letting code go live automatically." It resides on your machine 24/7, keeps applications running, and only sends notifications when human intervention is needed. It is officially said to provide all the advantages of PaaS, but without lock-in.

Purpose

Stakpak is aimed at DevOps work and can be used to generate infrastructure code, debug Kubernetes, configure CI/CD, and automate deployments, without handing production environment secrets directly to the LLM.

Main Capabilities

• Security mechanisms: Secret Substitution lets the LLM use credentials without seeing them; Warden Guardrails block destructive operations at the network layer; built-in DevOps Playbooks provide a curated DevOps knowledge base.
• Autopilot (24/7 autonomous runtime): Provides lifecycle aliases for one-click setup/start/stop, and also supports canonical subcommands. Before running on a remote VM, Docker must be installed and the current user must be able to access it; 2GB+ memory is recommended; Swap is strongly recommended for small Linux hosts; Linux user services may require linger to continue running after logout.stakpak up A preflight check is performed before startup,stakpak autopilot doctor can be used for deployment readiness checks.
• Configuration:~/.stakpak/config.toml controls profile behavior (model, allowed_tools, auto_approve, system_prompt, max_turns, provider credentials);~/.stakpak/autopilot.toml controls runtime wiring (schedules, channels, notification routes, service/server settings). The profiles of schedule and channel are independent of each other: schedule --profile monitoring controls the behavior of runs started by that schedule, channel --profile ops controls the session behavior of inbound messages from Slack/Telegram/Discord, and notification --target "#ops" only controls where notifications are sent. Notification routing uses channel (transport method, such as slack, telegram, discord) and target (destination within the transport, such as Slack #ops or C1234567890).
• Security hardening: mTLS end-to-end encrypted MCP; dynamic secret substitution; secure password generation; privacy mode can redact sensitive data such as IP addresses and AWS account IDs.
• DevOps work support: asynchronous task management (background commands such as port forwarding and servers, with tracking and cancellation); real-time progress streaming (real-time updates for long-running processes such as Docker builds and deployments); infrastructure code indexing (automatic local indexing and semantic search for Terraform, Kubernetes, Dockerfile, GitHub Actions); documentation research agent (built-in web search for technical documentation, cloud providers, and development frameworks); subagents (dedicated research agents for code exploration and sandbox analysis, enabled via --enable-subagents); batch message approval; reversible file operations (automatic backup and restore).
• Adaptive intelligence: Rule Books can customize agent behavior, with built-in standard operating procedures, playbooks, and organizational policies; persistent knowledge lets the agent learn from interactions and remember incidents, resources, and environment details.

Usage

Installation supports Linux, macOS, and Windows. Homebrew (Linux & macOS) can install and update it; platform-specific binaries can also be downloaded from GitHub Releases. Official images are provided that include commonly used CLI tools (docker, kubectl, aws cli, gcloud, azure cli, etc.).

CanUse your own Anthropic or OpenAI API key, a custom OpenAI-compatible endpoint, or a Stakpak API key.

• Option A: Use a Stakpak API key (no credit card required), run stakpak and follow the instructions to create a new key.
• Option B: Do not use a Stakpak API key, create ~/.stakpak/config.toml, and you can optionally use BYOK (use an Anthropic/OpenAI key), bring your own LLM (a local OpenAI-compatible endpoint such as Ollama or LM Studio), or mix built-in and custom providers.

Launch the TUI:stakpak; use Docker to launch the TUI.

MCP-related: can act as a secure MCP proxy or expose security-hardening tools through an MCP server. Tool modes include local (file operations and command execution only, no API key required), remote (AI-driven code generation and search tools, API key required), and combined (default, API key required). Additional flags: --disable-secret-redaction (not recommended, prints secrets in plaintext), --privacy-mode, --enable-slack-tools. The MCP proxy server can multiplex multiple upstream MCP servers via a configuration file. The configuration file discovery order is ~/.stakpak/mcp.toml or ~/.stakpak/mcp.json、.stakpak/mcp.toml or .stakpak/mcp.json、./mcp.toml or ./mcp.json. Servers can be added via the CLI or defined directly in ~/.stakpak/mcp.toml ; using disabled = true keeps the configuration but does not load it.

Agent Client Protocol (ACP): a standardized protocol that enables AI agents to integrate directly into code editors such as Zed, providing real-time AI chat, real-time code analysis, tool execution, session persistence, streaming responses, agent planning, and other features. You need to install Stakpak, configure the Zed editor (add it in ~/.config/zed/settings.json ), start the ACP agent, and then in Zed click Assistant (✨) → + → New stakpak thread to use it.

Rulebook management: use Stakpak Rulebooks to manage SOPs, manuals, and runbooks. Rulebooks are markdown files with YAML frontmatter.

Shell completion: supports tab completion for bash, elvish, fish, powershell, and zsh.

Shell mode: system commands can be explicitly executed from the input bar. See the Shell Mode documentation for details on background and foreground execution.

Platform testing: provides a comprehensive test report for Windows CLI functionality, including installation, configuration, and integration with WSL2 and Docker.

Applicable scenarios

Suitable for development and operations teams that need to automate DevOps processes, safely involve AI in production environment operations, and hope to reduce manual on-call work.

Limitations

The official documentation does not provide specific pricing, cost, or performance data; some features (such as sub-agents) require specific flags to enable; running on a remote VM has prerequisites such as Docker, memory, and Swap; when using a Stakpak API key, Brave users may encounter automatic redirection to a localhost port and need to manually copy the key.