LiteLLM Agent Platform is a self-hosted control plane that sits on top of agent runtimes such as OpenCode, OpenClaw, Hermes, Claude Managed Agents, the Cursor Agents API, and Deep Agents. You pick a runtime, create an agent, and give your team a single UI and API to run it, regardless of which runtime executes it underneath.
The platform manages access on behalf of developers, so they can create and run agents without Bedrock or Anthropic console credentials. It keeps persistent agent sessions across runs, retains memory so agents carry context between sessions, and runs agents on cron schedules. Agents can be wired to tools and skills from the creation screen and then dispatched to any registered runtime.
It runs from a Docker Compose stack that includes the web/API service and a Postgres database; optional profiles start template runtimes (OpenCode, Deep Agents, Hermes, OpenClaw) and register them in the UI automatically. Provider credentials are added in Settings before agents call hosted model providers. Agents are reachable through the UI, the API, webhooks, Slack, Microsoft Teams, and Google Chat.
Features
- Unified API: one interface to create and run agents across every supported runtime
- Supported runtimes: Claude Managed Agents, Cursor Agents API, OpenCode, OpenClaw, Deep Agents, Hermes Agent, plus Gemini Antigravity, Pydantic Deep Agents, and Elastic Agent Builder
- Access control: developers run agents without direct provider console access; sign-in via a master key
- Session management: persistent agent sessions that survive across runs
- Cron schedules: run agents on a timer
- Memory: agents keep context across sessions
- Tools and skills: attach tools and skills to an agent at creation time; MCP is supported
- Channels: UI, API, webhook, Slack, Microsoft Teams, and Google Chat
- Docker Compose profiles: start template runtimes and auto-register them as local-opencode, local-deepagents, local-hermes, or local-openclaw
