About CapaKit#
I'm building CapaKit because the way we write software has fundamentally changed, but our local dev environments haven't caught up.
When you let an AI agent write code, pull dependencies, and run scripts, you are delegating a massive amount of trust. You shouldn't have to choose between moving fast with coding agents and keeping basic control over your host machine.
Why CapaKit exists#
This is especially true for non-dev users of coding agents, but we sometimes give coding agents tasks and they are essentially creating "black boxes". Our way to monitor the work being done is to do what's called "black box testing" meaning, we see what goes in and what comes out.
Along the way, coding agents are making architectural choices, running npm install, and wiring up services.
But we are not really present when that is happening.
That changes the infrastructure problem.
CapaKit is my attempt to make agent-driven development safe and practical. It puts a hard boundary around the entire lifecycle, from build to run, so you can let agents do their job without handing them the keys to your filesystem and network.
Who I am#
I'm Roman Landenband, a software engineer based in Tel Aviv. Over the last 20 years, I've built backends, mobile apps, developer tools, and SaaS platforms.
I've been building with LLMs since the GPT-3 days. When Anthropic announced the MCP protocol, it was excited to see the
AI/LLM ecosystem finally standardizing.
I started hacking on what became CapaKit (originally called mcpgate.com) in early 2025 to solve the integration and
security friction I was seeing working with LLMs.
- GitHub: @romansky
- CapaKit org: github.com/capakit
- Contact: contact@capakit.com
Where this is going#
CapaKit is currently in public alpha. The runtime and CLI toolkit are free to use, and I intend to keep them that way.
We are in the incredibly early days of AI-native development. If you're building with agents, I'd genuinely love your feedback, bug reports, and especially security scrutiny.
— Roman Landenband