Stakpak is an open source agent that lets you hand day-to-day infrastructure operations over to AI. It runs as a Rust-based CLI tool on your own machine or server, staying resident in the background 24 hours a day and watching the state of your applications. When it detects a failure or anomaly, it tries to address it on the spot instead of waiting for a human to wake up. Expiring TLS certificates, idle cloud resources, connection errors — many of the things that drag operators out of bed at night are meant to be handled by the agent before they escalate. It is licensed under Apache 2.0, and is designed so that neither your code nor your data leaves your own environment.
A comparable idea is the PaaS (a platform such as Heroku or Render that takes over the tedious parts of operations for you), but a PaaS comes with a strong dependency on the vendor. Stakpak positions itself for people who want the ease of managed operations without giving up control of their infrastructure.
Key Features
- 24/7 autonomous operation (Autopilot): Runs continuously as a resident process, handling health checks, failure detection, and remediation without human involvement. Long-running work such as Docker builds and deployments is handled asynchronously with streaming progress
- Network sandbox (Warden): Acts as a transparent proxy that mediates traffic, with allowed actions defined in the Cedar policy language. Destructive operations and unauthorized network calls are blocked at the network level before they execute
- Secret substitution: Detects more than 210 types of secrets, including API keys and passwords, and replaces them with substitute values before anything reaches the model. The real values are restored only at the moment a command runs, so the model never sees your credentials
- Semantic search over infrastructure code: Automatically indexes Terraform, Kubernetes manifests, Dockerfiles, and GitHub Actions so it can answer questions like “where is this setting defined?”
- Rulebook (standard operating procedures): Lets you write your own procedures and playbooks to constrain how the agent behaves, aligning incident response with your team’s conventions
- Audit logs and rollback: Whole sessions are recorded and file changes are backed up automatically, so you can trace and reverse what the agent did
- BYOK (bring your own API key): Connect Anthropic, OpenAI, Gemini, or a local model endpoint of your choosing
Pricing
| Plan | Monthly price | Apps managed | Highlights |
|---|---|---|---|
| Open Source | $0 (free) | Unlimited (bring your own API key) | Fully open source, local execution, sandbox, secret substitution |
| Hacker | $15 | 3 | LLM gateway, private skills, watch triggers, centralized logging. Includes $10/month in credits |
| Builder | $45 | 10 | Everything in Hacker plus more credits ($35/month) and lower fees. Extra apps at $5 each |
| Team | $450 | 50 | Unlimited users, shared memory, admin dashboard, SSO, Slack integration, priority support |
| Enterprise | Contact sales | Unlimited | Self-hosting support, custom security policies, dedicated support |
Pricing is current as of August 2026. Please check the official site for the latest details.
The free Open Source plan is not a feature-limited tier: it is free in the sense that you supply your own model API key. The paid plans exist for the hosted LLM gateway and team-oriented features.
Pros & Cons
✅ Pros
- Apache 2.0 open source, running entirely on your own infrastructure, so you can use it on the premise that data stays in-house
- The sandbox and secret substitution are built in from the start, answering the obvious worries about letting AI touch production
- The free plan is not feature-limited, so you can evaluate it seriously with nothing more than your own API key
- A single Rust binary, installable via Homebrew, binary releases, or Docker
- Fits naturally alongside the tools teams already use: Terraform, Kubernetes, Docker, and the major cloud CLIs
⚠️ Cons
- Because the target is infrastructure operations, it is hard to use without background knowledge of Terraform or Kubernetes
- Requires Docker and 2GB or more of memory, so you need somewhere to run it
- Given that it remediates autonomously, pointing it at production before tightening your Rulebook and policies carries real risk
- Stakpak announced in July 2026 that it is joining Vercel, so you will want to follow official announcements about how the product and pricing evolve
Comparison with Similar Services
| Criteria | Stakpak | Terraform + manual operations | Heroku / Render (PaaS) | General-purpose coding agents |
|---|---|---|---|---|
| Primary role | Autonomous production operations | Infrastructure configuration management | Fully managed operations | Code generation and editing |
| Always running | Yes (24/7 resident) | No | Handled by the platform | No (per conversation) |
| Vendor lock-in | Low (runs in your environment) | Low | High | Medium |
| Production safety mechanisms | Sandbox, secret substitution, rollback | Designed by the operator | Left to the platform | Varies widely by implementation |
| Pricing model | Free to $450/month (by apps managed) | Tools are free; labor is the cost | Usage-based on resources | Mostly monthly subscriptions |
Who Is It For
- Small development teams running their own infrastructure who want fewer late-night alert responses
- Solo developers and startups who want PaaS-level convenience without depending on a single vendor
- People who want to delegate operations to AI but have hesitated over credential handling and destructive commands
- Engineers already using Terraform or Kubernetes who want to add a layer of automation on top
- Teams that want to evaluate a self-hostable open source agent internally
Summary
Stakpak takes the idea of “let AI run your infrastructure” and tries to shape it into something usable in production. Its defining traits are the two built-in safety mechanisms — the network sandbox and secret substitution — combined with 24/7 autonomous operation that never leaves your own environment. Since the free plan is not feature-limited, a realistic path is to bring your own API key, try it on a small test environment, build up your Rulebook, and only then widen its scope toward production. Note that Stakpak announced in July 2026 that it is joining Vercel, so it is worth checking the official site and the GitHub repository for the current state of things before adopting it.