Skip to main content

What is Roche?

Roche is a universal sandbox orchestrator for AI agents. It provides a single abstraction (create / exec / destroy) over multiple sandbox providers with AI-optimized security defaults — network disabled, filesystem readonly, timeout enforced.

Named after Édouard Roche — the Roche limit is the inviolable physical boundary for celestial bodies; Roche is the inviolable execution boundary for code.

The Problem​

Every AI agent framework independently integrates sandbox providers, creating an N×M complexity problem:

LangChain ──┐         ┌── Docker
CrewAI   ───┤  N × M  ├── E2B
AutoGen  ───┘         └── Modal

Roche reduces this to N+M:

LangChain ──┐              ┌── Docker
CrewAI   ───┤── Roche() ───├── Firecracker
AutoGen  ───┘              └── WASM

Features​

  • AI-safe defaults — network off, readonly filesystem, 300s timeout
  • Multi-provider — Docker, Firecracker, WASM behind a unified API
  • CLI + SDKs — roche binary + Python & TypeScript SDKs
  • Framework-agnostic — works with LangChain, CrewAI, AutoGen, OpenAI Agents SDK, Anthropic API, Camel-AI
  • Zero config — sensible defaults, opt-in for permissions

Quick Example​

from roche_sandbox import Roche

roche = Roche()

with roche.create(image="python:3.12-slim") as sandbox:
    output = sandbox.exec(["python3", "-c", "print('Hello from Roche!')"])
    print(output.stdout)  # Hello from Roche!
# sandbox auto-destroyed

Relationship to Castor​

Castor and Roche are orthogonal — they solve different layers of the AI agent security stack:

LayerToolPurpose
LogicalCastorCapability budgets, HITL approval, checkpoint/replay
PhysicalRocheProcess/container/VM isolation

Operators can use both, either, or neither. Castor manages what tools an agent can call. Roche manages what system resources that code can access.

Next Steps​