Motus is an open-source agent serving project designed to enable higher capability, lower cost, and faster AI agents. It provides the necessary infrastructure for efficient agent serving, abstracting away deployment complexities for self-managed and cloud environments at any scale. The project targets developers building and deploying AI agents, offering a unified approach that works seamlessly with various coding agents and SDKs.

How It Works

Motus employs a serving runtime that automatically transforms Python code into parallel, resilient workflows. Its core innovation lies in using @agent_task decorators to define asynchronous tasks, allowing the runtime to infer dependency graphs dynamically without explicit DAG definitions. This approach facilitates automatic parallelism, scheduling, caching, and observability, simplifying the development of complex agent behaviors. It supports a multi-provider model architecture and integrates smoothly with existing agent SDKs.

Quick Start & Requirements

Installation is straightforward via a shell script (curl -fsSL https://www.lithosai.com/motus/install.sh | sh) for the CLI and plugin, or using package managers (uv add lithosai-motus or pip install lithosai-motus) for the library. Motus is designed to work out-of-the-box with coding agents like Claude Code, Codex, and Cursor. It supports agents built with the OpenAI Agents SDK, Anthropic SDK, Google ADK, and plain Python. Official documentation, quickstart guides, and examples are available via links provided by LithosAI.

Highlighted Details

• Agent Core: Features a ReActAgent with built-in multi-turn memory, structured output via Pydantic, and input/output guardrails, enabling agents in under 10 lines of code.
• Flexible Tooling: Exposes Python functions as tools (@tool), supports class methods, MCP servers, nested agents (agent.as_tool()), and untrusted code execution via Docker sandboxes.
• Task-Graph Runtime: @agent_task decorators convert functions into nodes in a dependency graph with automatic parallel execution, retries, timeouts, and multi-return futures.
• Observability: Integrated tracing for LLM calls, tool invocations, and task dependencies, with an interactive HTML viewer and Jaeger export options.
• Multi-Provider Models: Unified client supports OpenAI, Anthropic, Gemini, and OpenRouter, with seamless switching and local model integration (Ollama, vLLM).
• SDK Compatibility: Offers drop-in replacements for OpenAI Agents SDK, Anthropic SDK, and Google ADK.
• Human-in-the-Loop: Built-in support for interactive agent execution, allowing human approval or clarification during runtime.

Maintenance & Community

The project encourages community involvement via a Slack channel and provides a detailed Contributing Guide.

Licensing & Compatibility

Motus is licensed under the Apache 2.0 license, which generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

The provided README does not explicitly detail limitations such as alpha status, known bugs, or unsupported platforms. The project appears focused on providing a robust serving infrastructure.