Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.litellm-agent-platform.ai/llms.txt

Use this file to discover all available pages before exploring further.

LiteLLM Agent Platform (LAP) is self-hosted infrastructure for running coding agents inside isolated Kubernetes sandboxes. Each sandbox gets a credential vault that replaces your real API keys and tokens with stub placeholders — the agent process can never read the real values, because the vault sidecar swaps them for real credentials at the wire level on every outbound TLS connection. You interact with sandboxes through the lap CLI, the web UI, or the REST API.

Key features

Isolated sandboxes

Every session runs in its own Kubernetes pod. Agents are fully isolated from each other and from the host environment.

Credential vault

Agents receive stub tokens (e.g. GITHUB_TOKEN=stub_github_a8f1). The vault sidecar intercepts outbound HTTPS and swaps stubs for real keys before the request leaves the pod.

Terminal-first CLI

The lap CLI attaches your local terminal to a sandbox over WebSocket — same feel as SSH, but sandboxed. SIGWINCH is forwarded so your window size tracks perfectly.

Multiple harness types

Run Claude Code, Codex (TUI harnesses) or opencode, claude-agent-sdk (API harnesses) — each in its own pre-configured container image.

Warm pool

Pre-provisioned pods wait in the cluster so session creates are near-instant. With a warm pool, a new sandbox is ready in under two seconds.

Web UI and REST API

Create agents, manage sessions, and send messages through the browser or directly with curl. Full OpenAPI-compatible REST API included.

The vault proxy

Every sandbox pod runs a vault sidecar that proxies all HTTPS egress. When the agent makes an outbound TLS connection, the sidecar intercepts it and replaces any stub credential in the request with the real value — before the bytes leave the node. The agent process sees only stubs:
GITHUB_TOKEN=stub_github_a8f1
LITELLM_API_KEY=stub_litellm_bb20
The wire carries real values. The agent can echo $GITHUB_TOKEN and only ever gets the stub back — the real token is never in the process environment.

Available harness types

HarnessTypeDescription
claude-codeTUIClaude Code CLI under node-pty, attached to your terminal via WebSocket
codexTUIOpenAI Codex CLI, same PTY-over-WebSocket model
opencodeAPIopencode in server mode; interact via POST /sessions/{id}/message
claude-agent-sdkAPIAnthropic Claude Agent SDK; task-based, no TTY required
TUI harnesses are designed for interactive terminal use with lap. API harnesses accept structured messages and return structured replies, making them well-suited for automation and programmatic workflows.

Next steps

Quickstart

Install the lap CLI and open your first sandbox in minutes.

Self-hosting

Deploy LAP on Kubernetes locally with kind, or on AWS EKS for production.