Fred K. Schott pivoted from Astro and built a sandbox framework for AI agents.
The project is called flue, written in TypeScript, Apache-2.0 licensed, and just released version 0.4.1. 220 commits, 9 branches — not idle.
What Problem Does It Solve
Agents can write code now, but where do they run is the real problem.
You cannot let LLM-generated code execute directly on your production servers. Claude Code, Codex, Cursor — these tools all run in their own sandboxes. But if you want to build your own agent application, the sandbox layer is on you: container management, permission control, filesystem isolation, network restrictions. Each layer is nontrivial.
flue abstracts this layer out. Its positioning is clear: provides the scheduling layer for agents, including sandbox environments, skill management, and connectors.
From the README and code structure, the core components include:
- packages:
@flue/sdkand@flue/clitwo core packages - connectors: connectors for various external services
- examples: runnable examples
- AGENTS.md: Agent instruction file explaining how to develop the project with AI tools
Why It Is Worth Watching
Not because of star count — currently 3K, not eye-catching among AI frameworks. What makes it worth watching is who is building it.
Fred K. Schott is the co-founder of Astro. What Astro did in the static site generator space is "islands architecture" — only hydrate the parts that need interaction. This "execute on demand, minimum privilege" philosophy is highly consistent with the security philosophy of agent sandboxes.
flue is like porting Astro's design philosophy into the agent world: agents should execute in a constrained, predictable environment, accessing external resources only when needed.
Comparison with Similar Projects
This space is getting crowded:
- Google ADK (Agent Development Kit) leans more toward Google ecosystem integration
- OpenClaw, Hermes Agent are full agent runtime environments
- Manus (acquired by Meta) went the cloud agent product route
flue's differentiation: it only does the sandbox and scheduling layer, without locking you into any model or agent framework. You can hook it behind Claude Code, or behind OpenAI's Agent SDK.
Can You Use It Now
Still early. Version 0.4.1 means the API may still change. But the code quality shows a complete monorepo structure, issue picker config, CI/CD pipeline — this does not look like an experimental project.
Who it is for: teams wanting to build their own agent execution environment, especially TypeScript stack developers. If you just need a sandbox to run agent-generated code, flue is much easier than building Docker isolation from scratch.
Who it is not for: anyone wanting an out-of-the-box complete agent product. flue is infrastructure, not the application layer.
The next thing to watch is whether its connector ecosystem grows. The value of a sandbox framework is not the sandbox itself — it is how many external services it can connect to.
Primary sources:
- GitHub: https://github.com/withastro/flue
- Project README and code structure analysis