MS-Agent is a lightweight, open-source agent framework from Alibaba’s ModelScope team, built for complex, long-running tasks. It combines models, tools, skills and sub-agents under a customizable harness that manages planning, context, permissions and execution feedback.
The framework separates session history, active context and project memory into distinct layers, with context compression that prunes tool outputs and summarizes older context so long tasks stay within model limits. A lead agent can delegate open-ended work to specialist sub-agents, or you can define explicit sequential and DAG workflows with per-role models, tools and skills. A shared Python SDK powers the CLI, terminal UI and WebUI, so the same agent logic runs everywhere, and scheduled (cron) jobs support recurring research and monitoring. MS-Agent ships with substantial built-in applications: Agentic Insight for deep research, CodeGenesis for generating software projects from requirements, FinResearch for financial research, DocResearch for turning source documents into illustrated reports, and Singularity Cinema for short-video generation. The framework grew out of the earlier ModelScope-Agent project, and its Agent Hub can manage agent workspaces across several popular agent frameworks, not just its own. It is Apache-2.0 licensed, has more than 4,400 stars on GitHub, and was under active development in 2026.
Free. MS-Agent is open source under the Apache-2.0 license; you supply your own model API keys or local models.
Developers building serious, long-running agent applications — especially deep-research, document and code-generation workflows — who want memory management, workflow control and ready-made application templates in one framework.