Hey! We’re Jacob Wright & Cambree Bernkopf, founders of Castari.
TL;DR
Castari lets teams deploy AI agents to secure, autoscaling sandboxes in seconds.
You write agent logic with the Claude Agent SDK (other frameworks coming soon). We turn it into a production-grade endpoint.
We’re especially focused on:
- Teams building internal agents (ops, support, data, eng workflows).
- Companies whose core product is an agent built on Claude Agent SDK and need a safe, reliable way to run it.
Launch Video:
The Problem:
Serious agents are not just a single API call. Agents need a computer.
They execute code, call tools and browsers, run for minutes instead of milliseconds, and touch sensitive systems. Most teams end up rebuilding the same fragile setup:
- Homegrown agent harnesses
- Sandboxes, file systems, and queues stitched together
- Logs and traces scattered across many services
- Ad hoc snapshots and risky secret handling
The result:
- Every agent needs its own mini runtime
- Debugging one agent run feels like detective work
- Adding a new agent feels like starting a new infra project
Good agents stall at “cool demo” instead of reaching production. Deploying agents becomes a full-time infrastructure problem.
How Castari solves it
We’re building Castari to be the natural runtime for AI agents built on Claude Agent SDK. We wrap the entire agent in a sandbox, deploy it, and provide you with an endpoint url that auto-scales as requests come in. No need to write code in your tools to manage your sandbox lifecycle.
1. Drop in a config file → get a production runtime
In your Claude Agent SDK repo:
- Add a
castari.yaml(entrypoint, tools, env vars) - Run
castari deploy
That’s it. Your agent is now running in a secure sandbox with an endpoint, UI, and observability.
Whether that agent is:
- an internal system your team uses
- a customer-facing agent that is your product
the deployment workflow is identical and boring (in the good way).
2. Sandboxing & safety, built-in
- Every run executes in an isolated sandbox designed for tool-using agents
- You design the behavior; we handle isolation and scaling
- No more duct-taping together sandbox primitives for specific tools
Bottom line: You keep building on Claude Agent SDK the way you already are — for internal workflows or core products. Castari turns that into a safe, observable, production-ready runtime.
Why Castari is different
E2B, Modal, Daytona, Cloudflare (and friends) give you powerful sandbox and compute primitives.
Castari sits one layer up:
Castari wraps around your Claude Agent SDK-based agents so you can define them declaratively and run them in secure sandboxes, without owning the underlying sandbox lifecycle.
- Agent-first semantics: runs, sessions, tools, snapshots are native concepts we manage.
- Zero framework lock-in: remove Castari and your agent still runs on vanilla Claude Agent SDK; keep Castari and your infra team gets its life back.
- Focused start: deeply integrated with Claude Agent SDK today, with the same “drop-in config” experience coming to other frameworks next.
We’re building the runtime we wish existed for agents.
Why we’re working on this
While I was leading the FDE team at RunPod:
- Built and shipped AI infrastructure and agents for Fortune 500s and fast-growing startups
- Watched the same loop repeat: agent works in a notebook → weeks disappear into infra glue → nobody trusts the system, production release gets delayed (over and over again).
Additionally, while @Cambree Bernkopf and I scaled our AI consumer app to 2m+ members, our biggest issues were getting our agents to work reliably in production.
The pattern is clear: agents deserve a first-class runtime, the way frontends got one.
Castari is our bet that:
- Sandboxing & safety should be built-in, not bolted on
- Going from “prototype” to “production-safe” should be one config file, not a quarter-long project
- AI teams should ship agents, not infrastructure.
Ask: early pilot access
We are opening a limited pilot for:
- Teams building internal agents who need a safer, more reliable runtime
- Companies building agent centric products that need to run in production without bespoke infra
If that is you and you are tired of wrestling with sandboxes, observability, and brittle tooling:
→ Email jacob@castari.com with subject “Castari Pilot” and 2 to 3 sentences about your agent stack and use case.
→ Check out our open source repo for using any model with Claude Agent SDK on GitHub.
→ Join the waitlist on castari.com to hear about our first releases.
Jacob, Cambree, and the Castari team