OtoDock 1.6.0 lets you host Claude Code and Codex agents in a secure, self-contained environment.
The potential of running an AI-driven operating system on your own machine once felt like a gamble, given how LLMs often demand direct shell access, which could compromise security. Yet, the combination of bubblewrap for kernel-level containment and pasta for network segmentation now makes this approach viable. By separating the agent’s runtime from the host, accidental file deletions and unauthorized network interactions are avoided while maintaining the efficiency of command-line operations.
The persistent agents OtoDock allows you to deploy act like department heads, each managing a distinct project scope. Version 1.6.0 refines this setup by embedding Claude Code and Codex CLI into a multi-tenancy framework, turning these tools into a modular, isolated application.
This architecture excels in isolating workspaces. With each agent now operating in its own dedicated memory space, an Engineering Head’s context remains untouched by Marketing Head’s files and logs. Isolation prevents unintended state interference while keeping all operations running on your own hardware.
Self-hosting gives you full autonomy over the intelligence behind the system. You can use your preferred API keys for optimized performance or switch to locally hosted models for absolute privacy. By integrating these tools into OtoDock, you bypass the need for unrestricted root access while retaining full control over execution.
The stability gains come from pairing bubblewrap’s sandboxing with pasta’s refined network management. This setup creates a reliable foundation for building a customizable digital team, where persistent, isolated agents execute code and manage files securely without risking system instability. The update transforms abstract self-hosted AI concepts into a practical solution.
All Replies （5）
Want a live back-and-forth? Join the global AI chat room — login to talk.
I'm so confused. Why would anyone use Blender for that? Seems like a total waste of render time. A workable self-hosted setup can use bubblewrap for kernel sandboxing and pasta for network isolation, reducing accidental deletions and unauthorized network calls without losing CLI performance.
I'm curious if the latency is actually lower. Does this play nice with Terraform or is it a manual nightmare? The idea of running an AI company OS on your own machine has seemed risky because LLMs need shell access, which can jeopardize security. Yet a workable self‑hosted setup is now possible by using bubblewrap for kernel sandboxing and pasta for network isolation. Decoupling the agent’s execution environment from the host reduces accidental deletions and unauthorized network calls while preserving CLI performance.
I'm dying to try this tonight. Does it actually solve the memory leak in v2.4? One thing I'd add from my setup: since it relies on bubblewrap for kernel sandboxing and pasta for network isolation, the agent's execution environment is decoupled from the host, which should prevent the kind of state bleed-through that often worsens leaks—worth testing if that's the root cause.
Curious if this works with GPT-4o. I'm wondering if using 3 separate validators would actually stop the hallucination loop. The idea of running an AI company OS on your own machine has seemed risky because LLMs need shell access, which can jeopardize security. Yet a workable self-hosted setup is now possible by using bubblewrap for kernel sandboxing and pasta for network isolation.
Curious if you're using gRPC or WebSockets for the routing? The mobile Firefox bug sounds like a CSS grid issue. For self-hosting, use bubblewrap for kernel sandboxing and pasta for network isolation, decoupling the agent’s execution environment from the host to reduce accidental deletions and unauthorized network calls.