lithos-ai/motus

What it solves

Motus is an open-source agent serving infrastructure that simplifies the deployment and scaling of AI agents. It removes the need for complex frameworks by providing a runtime that handles the heavy lifting of scheduling, parallelism, caching, and resilience, allowing developers to build agents using plain Python code.

How it works

Motus uses a serving runtime that automatically converts Python functions (marked with @agent_task decorators) into a parallel dependency graph. This allows the system to infer data flow and execute tasks in parallel without requiring the developer to write explicit Directed Acyclic Graphs (DAGs). It supports a wide range of agents built with various SDKs (OpenAI, Anthropic, Google ADK) or plain Python, and provides a single command to serve agents locally or deploy them to the cloud.

Who it’s for

Developers building AI agents who want a lightweight, framework-less approach to serving and deploying their agents at scale, as well as those using coding agents (like Cursor or Claude Code) to accelerate their development workflow.

Highlights

• Task-Graph Runtime: Automatically handles parallel execution, retries, and timeouts for Python functions.
• Coded-based Workflows: No explicit DAGs required; dependencies are inferred from data flow.
• Multi-Provider Support: Unified client for OpenAI, Anthropic, Gemini, and OpenRouter, including support for local models via Ollama or vLLM.
• Multi-Agent Composition: Ability to wrap agents as tools (agent.as_tool()) for supervisor agents to coordinate.
• Integrated Observability: Automatic tracing of LLM calls and tool invocations with an interactive HTML viewer or Jaeger export.
• Docker Sandboxes: Capability to run untrusted code in isolated containers as tools.
• MCP Integration: Support for Model Context Protocol (MCP) compatible servers.
• Human-in-the-loop: Built-in support for interactive approval and feedback during execution.