ModelScope-Agent
also known as MS-Agent
Type: full-code · Vendor: Alibaba ModelScope · Language: Python · License: Apache-2.0 · Status: active · Status in practice: emerging · First released: 2023-08-07
Provide a lightweight, extensible Chinese-ecosystem agent framework with RolePlay agents, tool calling, hybrid RAG, and MCP-mediated multi-agent workflows.
Description. ModelScope-Agent (now also branded MS-Agent) is Alibaba ModelScope's open-source agent framework. It started in August 2023 as a RolePlay-style agent with tool calling and a published preprint (arXiv:2309.00986), and has since added an AgentFabric UI for creating agents tailored to applications, hybrid search combining FAISS dense retrieval with BM25, MCP-based tool calling, and code-generation artifacts. The framework integrates Qwen and other ModelScope-hosted models and ships RAG over LlamaIndex.
Agent loop shape. RolePlay-style tool-calling agent. Agents are configured with a role, tools, and an LLM (typically Qwen or OpenAI-SDK-compatible). The loop is the standard reason-tool-observe pattern with MCP servers and local tools as the action space; a multi-agent path orchestrates specialised agents for task decomposition.
Primary use cases
- Chinese-ecosystem tool-using agents over Qwen and ModelScope APIs
- agent chat with MCP-mediated external tools
- hybrid (BM25 + dense) RAG over local code and docs
- user-built agents via the AgentFabric UI
Key concepts
- RolePlay agent (docs) — Primary agent class — a role-prompted, tool-using agent supporting image inputs (e.g. GPT-4o) and OpenAI-SDK tool calling.
- AgentFabric (docs) — Interactive UI for creating agents tailored to applications, released Nov 17 2023.
- MCP-based tool calling → mcp (docs) — Agent chat with tool-calling capabilities based on the Model Context Protocol.
- Hybrid search → hybrid-search (docs) — FAISS dense retrieval combined with BM25 sparse retrieval for RAG.
- Multi-Agent (general purpose) → orchestrator-workers (docs) — Orchestrates multiple specialised agents for task decomposition, data collection, quantitative analysis.
Patterns this full-code implements —
- ★★Tool Use
RolePlay agent supports OpenAI-SDK-style tool calling on Qwen and other backends; MCP servers extend the tool surface.
- ★★ReAct
LLMAgent.run_loop is documented as a llm-calling and tool-calling loop with per-step think→tool-call→tool-result iteration; the framework's autonomous exploration is built on this iterative tool-usin…
- ★★Model Context Protocol
MCP-based tool calling is highlighted on the README; the agent connects to MCP servers for external tools.
- ★★Hybrid Search
Hybrid Search is named as a feature: FAISS dense retrieval combined with BM25 sparse retrieval; the Skill Module also uses it for skill retrieval.
- ★★Agentic RAG
RAG flow is built on LlamaIndex with hybrid retrieval; more recent versions add Sirchmunk-based knowledge search over local codebases and documentation invoked during agent conversations.
- ★Code-as-Action Agent
Agents generate and emit code as artifacts; Code Genesis is the multi-agent code-generation project that produces production-ready software projects from natural language requirements.
- ★★Orchestrator-Workers
The Multi-Agent path orchestrates specialised agents; FinResearch is the concrete five-agent DAG: Orchestrator decomposes tasks, Searcher/Collector gather data, Analyst computes, Aggregator finalises.
- ★Skill Library
The MS-Agent Skill Module is an explicit implementation of Anthropic's Agent Skills protocol with hybrid skill retrieval, DAG-based execution, and progressive skill analysis. AgentFabric is a separat…
- ★★Structured Output
Tool-call schemas via OpenAI SDK provide a structured channel; Agentic Insight v2 also persists structured artifacts to disk as a file-system-as-context substrate. No first-class structured-response_…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Listed as alternative by (1)
References
“MS-Agent is a lightweight framework designed to empower agents with autonomous exploration capabilities.”
ModelScope-Agent preprint
“Sep 2, 2023: The preprint paper associated with this project was published.”
“Aug 7, 2023: The initial version of the modelscope-agent repository was released.”
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified