The team behind LiteLLM has published a platform designed to run fleets of coding agents inside isolated Kubernetes sandboxes. The system aims to provide a secure, scalable environment for developers and teams without exposing real credentials to agent code.
How it works
- Each agent session runs in its own fresh Kubernetes pod; the pod persists for 24 hours after detaching.
- A credential vault ensures agents never see real API keys: agents are presented only stub tokens.
- On outbound network connections the platform transparently substitutes the real secrets, preventing a rogue or buggy agent from leaking actual API keys.
This approach is particularly useful when scaling from a single coding agent to multiple agents used by a team, as it reduces credential exposure risk.
Supported agents and sandbox features
The platform provides agent-agnostic sandboxes with first-class support for:
- Claude Code
- Codex
- Hermes
All supported agents run in isolated pods subject to the 24-hour persistence window after detachment.
Workflows and developer API
There are two primary interfaces for users:
- Terminal-first CLI: open a sandbox, attach your terminal, perform work, and detach with Ctrl-D. This delivers a simple, terminal-centric workflow.
- Full developer REST API: create agents, open sessions, send messages, and read replies via REST calls, enabling custom orchestration on top of the platform.
Licensing and deployment
The project is open source under the MIT license, includes local development support, and provides a deployment path for production on AWS. The repository is available at github.com/BerriAI/litellm-agent-platform.
Summary
LiteLLM’s platform emphasizes security and scalability through isolated Kubernetes pods, a credential vault that protects API keys, and convenient CLI and REST interfaces for developers. These features make it easier and safer to scale coding agent usage across teams.
