Flue: The Sandbox Agent Framework from Astro (Fast Look)

Most agent frameworks hand you an SDK and call it done. So good luck wiring up sessions, error recovery, and state persistence yourself.

Flue is different. It’s a full agent harness — sessions, tools, skills, sandboxed execution, and durable recovery — all in one TypeScript framework from the Astro team. 5,935 stars in two days. And that’s not hype — that’s signal.

What Problem Does Flue Actually Solve?

If you’ve built an AI agent, you know the pattern: pick an LLM provider, write a loop, add some tool calls, pray it doesn’t crash mid-conversation. But most “agent frameworks” are just SDKs with a README — you’re still on your own for runtime, state, and error handling.

Flue flips that. It gives you the entire runtime from day one:

I ran npm create flue@latest my-agent on my MacBook Air M3 — scaffolded in about 30 seconds. And the generated agents/triage.ts had a working agent config with tool registration, skill hooks, and sandbox settings. And wired in my OpenAI key — it just ran. No Docker, no Docker Compose file, no 15-minute “getting started” doc crawl.

One thing I noticed testing the scaffold: the sandbox config was already set to local by default. I didn’t have to hunt through docs to figure out how isolation works — it’s on by default. That’s a level of polish most agent frameworks don’t ship on day zero.

How Flue Stacks Up

Because this isn’t another DeerFlow wrapper or a PilotDeck orchestration layer. Flue sits in a different category — the agent harness tier. Here’s how the field breaks down:

What jumps out? Flue’s durable execution and MCP support are unique right now. Yet no other framework — not even options like sandboxd — ships both. Still, DeerFlow has the star count — but it’s fundamentally a LiteLLM wrapper, different use case entirely.

The Catch with Flue (There’s Always One)

But Flue is also two days old. The docs are solid — Astro team quality, no surprise there. But the plugin ecosystem is empty, there’s no community yet, and some edge cases around the remote sandbox provider aren’t fully documented. If you need a battle-tested production agent framework for a client deliverable next week, this isn’t it.

Still, what Flue is good for: prototyping, personal agent projects, and getting ahead of the curve. The architecture is right — durable execution + sandbox + skills is the winning formula. And it just needs time to mature. Want to run a 24/7 agent? A cheap VPS droplet is enough for Flue’s Node.js runtime — no heavy infrastructure required.

Bottom Line on Flue

Flue is the most complete “plug and play” agent harness I’ve seen land this year. If you’ve been stitching agents together from SDK snippets and shell scripts, this is worth your evening. One command to scaffold, one file to configure, and you’re running. At 5,935 stars and climbing, the Astro team is onto something real.