MS-Agent
An MCP-enabled framework for research, code, document, finance, and video workflows.
- Source repo
- modelscope/ms-agent
- Stars
- ★ 4.4k
- Last updated
- 5d ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 67/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Python developer who needs asynchronous chat with tools exposed by a self-hosted or ModelScope MCP endpoint can build on
LLMAgent. - Not a fit if
- Users unwilling to install Node.js 22, pnpm, and uv for the WebUI build
- Projects still importing the legacy modelscope-agent package and paths
- Source review
- 67/100 · Some gaps
What does this agent do, and when should you use it?
MS-Agent is a lightweight Python framework for autonomous exploration and complex task execution. Its documented entry points include MCP-configured LLMAgent, AgentLoader, and the ms-agent CLI for project workflows. The repository packages Agentic Insight deep research, Doc Research, Code Genesis, FinResearch, and Singularity Cinema, with a React/FastAPI WebUI using real-time WebSocket communication. Its Agent Skills module implements the Anthropic Agent Skills protocol with hybrid retrieval, dependency-aware DAG execution, progressive loading, and controlled execution. Depending on the workflow, it produces research reports, document-analysis results, software projects, or short videos and may require model, search, or Docker services.
After installation, developers can configure a streamable HTTP MCP server in mcpServers, instantiate LLMAgent(mcp_config=mcp), and call await llm_agent.run(...). Agentic Insight v2 assigns research work to a Researcher and tool-augmented Searcher/Reporter sub-agents, persists structured artifacts to disk, and writes reports from an indexed evidence base. Doc Research accepts multiple files or URLs, analyzes them, and produces text-and-image Markdown reports; its release notes also list HTML, PDF, PPTX, and DOCX export. Code Genesis runs through ms_agent/cli/cli.py run --config projects/code_genesis, using either a seven-agent or four-agent pipeline to move from requirements through implementation and refinement, with code written to output. FinResearch uses DagWorkflow to coordinate collection, analysis, and aggregation, while processing data in an isolated Docker sandbox.
- A Python developer who needs asynchronous chat with tools exposed by a self-hosted or ModelScope MCP endpoint can build on
LLMAgent. - A research team that needs evidence-grounded reports, disk-persisted artifacts, and concurrent specialist agents can run Agentic Insight v2.
- An engineering team turning a natural-language brief into a static homepage or fuller software project can use the standard or simplified Code Genesis workflow.
- An analyst investigating uploaded files and URLs can use Doc Research to produce a Markdown research report.
- A market researcher covering mainland China, Hong Kong, or U.S. equities can use FinResearch for financial data, news/sentiment research, and report generation.
- A content team that wants a short video from a sentence or knowledge document can use Singularity Cinema for scripting, storyboards, voiceover, charts, subtitles, and generation.
How do you install or deploy this agent?
For base functionality:
pip install ms-agentFor deep-research functionality:
pip install 'ms-agent[research]'Or install from source:
git clone https://github.com/modelscope/ms-agent.git
cd ms-agent
pip install -e .The README specifies Python >=3.10. The default ModelScope inference path requires MODELSCOPE_API_KEY; the memory example additionally requires mem0ai and DASHSCOPE_API_KEY.
How do you use this agent?
Set credentials for the default inference path:
export MODELSCOPE_API_KEY={your_modelscope_api_key}Create an MCP configuration with mcpServers, instantiate LLMAgent(mcp_config=mcp), and invoke await llm_agent.run('Introduce modelscope.cn') from asyncio.run(main()). To launch the WebUI:
ms-agent uiIt opens at http://localhost:7860 by default. A first Code Genesis invocation is:
PYTHONPATH=. openai_api_key=your-api-key openai_base_url=your-api-url python ms_agent/cli/cli.py run --config projects/code_genesis --query 'Build a static site to display skills, projects, and contact info' --trust_remote_code trueWhat are this agent's strengths and limitations?
- It combines MCP chat, deep research, document research, code generation, finance research, and video generation as documented projects in one repository.
- Agentic Insight v2 uses the filesystem for context handoff and an indexed evidence base for report writing, which suits resumable, long-running research.
- Agent Skills combines FAISS and BM25 retrieval with skill dependency DAGs and progressive resource loading instead of loading all skill material at once.
- Code Genesis documents both a seven-agent production pipeline and a four-agent prototype pipeline, plus LSP-integrated validation and automated refinement.
- The React/FastAPI WebUI with WebSockets provides a documented local interactive surface.
- The default inference path depends on ModelScope API access and
MODELSCOPE_API_KEY; the memory example additionally relies on DashScope because ModelScope API-Inference does not yet provide embeddings. - Workflow requirements vary: Code Genesis, FinResearch, and Singularity Cinema use additional OpenAI, search, or image-generation-related environment variables.
- Isolated execution for Skills and FinResearch depends on Docker/ms-enclave, adding local deployment and security operations requirements.
- Some examples require
--trust_remote_code true, so adopters must assess the remote-code trust boundary. - Versions through v0.8.0 used the
modelscope-agentpackage andmodelscope_agentimports, creating a migration consideration for older integrations.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MS-Agent This agent | 67 · Some gaps | Web appFree + model costs | ★ 4.4k | 5d ago | Python | — |
| Codemini CLI | 69 · Some gaps | CLIFree + model costs | ★ 31 | 4d ago | JavaScript | OpenAI API · Claude API |
| Locally Uncensored | 72 · Some gaps | Desktop appFreemium | ★ 3.6k | 2d ago | TypeScript | OpenAI API · Claude API |
| Code Puppy | 56 · Major gaps | CLIFree + model costs | ★ 878 | 2d ago | Python | Codex · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The materials show permission checks, WebUI authentication, upload/download dry runs, backups, and restoration. However, CLI uploads default to public visibility, watch can continuously sync, and the sample does not fully specify authorization prompts or least-privilege policy, so this does not merit a high score. Key setup is documented, but evidence for sensitive-data storage, log redaction, and dependency vulnerability management is limited. License and copyright attribution are clear.
The README, CLI, WebUI backend, and tests present a broadly consistent framework structure; installation notes list Python, Node, pnpm, and uv requirements and explain differences between PyPI and source releases. Error handling and recovery mechanisms are visible, but static materials cannot establish runtime behavior, and CI uses a self-hosted runner, so dependency availability and failure feedback do not receive full marks.
Documentation covers CLI, TUI, WebUI, Python SDK, research, code generation, video, and several integration paths, with environment requirements for the interfaces. Scenario and environment coverage is strong. Permissions, execution boundaries, and skill triggers are described in part, but the sample does not systematically define capability limits or trigger conditions.
The repository separates SDK, configuration, tools, interfaces, projects, and documentation; it provides detailed installation steps, examples, bilingual documentation entry points, and a clear Apache-2.0 license. The README has release notes across several periods and explains source-versus-package differences, but the sample does not clearly identify maintenance ownership or provide a complete support and compatibility policy.
The framework combines models, tools, skills, workflows, and multiple interfaces, with a reusable SDK, domain applications, and Agent Hub, showing substantial extension value. Examples support a quick start, but installation and extensions require several runtimes and network dependencies; the cost-benefit of complex long-running tasks is not quantified in the evidence.
README installation and feature descriptions are partly corroborated by CLI arguments, the FastAPI backend, and test examples; the repository also includes explicit license, version, and configuration material. Many performance, leaderboard, and capability statements remain README claims, and the sample contains only limited test excerpts. This review distinguishes documented claims from visible source facts and does not infer execution results.
- CLI uploads default to creating a public repository; before uploading or enabling continuous sync, check repository visibility, directory scope, and credentials.
- First launch may install backend and frontend dependencies over the network; the source tree and current PyPI release differ in interface and extension capabilities.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/modelscope-agent)<a href="https://followagents.com/en/agents/modelscope-agent"><img src="https://followagents.com/badge/modelscope-agent.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
🌐 English Version Release v1.6.0 🚀 New Features & Enhancements - Agentic Insight v2 Enhancement: Architecture and performance optimization for the deep research system. ([#886](https://github.com/modelscope/ms-agent/pull/886)) - Knowledge Search Support: Added Sirchmunk-based knowledge search module, enabling automatic retrieval over local codebases and documents during agent conversations. ([#882](https://github.com/modelscope/ms-agent/pull/882)) …View on GitHub ↗ -
Release notes excerpt
🌐 English Version Release v1.6.0rc1 🚀 New Features & Enhancements * DeepResearch Refactor: Significant refactoring of the DeepResearch module for better performance and structure. ([[#858](https://github.com/modelscope/ms-agent/pull/858)](https://github.com/modelscope/ms-agent/pull/858)) * Video Generation Improvements: …View on GitHub ↗ -
Release notes excerpt
🚀 Release Notes - v1.6.0rc0 The v1.6.0rc0 version introduces significant enhancements to the WebUI, multi-modal capabilities (video and animation), and a major refactor of the Skills system. 🌟 Highlights * WebUI: A new, simple WebUI has been added, featuring "Code Genesis," video support, and enhanced project management. * Skills System Refactor: Multi-skills DAGs analysis and execution; New skill pipeline. * Multi-modal Enhancements: Introduction of a new animation engine and major refactors to video generation and subtitle processing. --- 📝 What's Changed …View on GitHub ↗
FAQ
Is ModelScope required for basic agent chat?
LLMAgent uses ModelScope's API inference service by default and requires MODELSCOPE_API_KEY. Some project examples use OpenAI API credentials and base URLs, and the release notes also state support for Anthropic API format.Does Agent Skills execute skill code directly?
What else is needed for FinResearch?
OPENAI_API_KEY and OPENAI_BASE_URL, the full workflow needs either EXA_API_KEY or SERPAPI_API_KEY. The README also shows a reduced DagWorkflow configuration that omits the search engine.Where are generated outputs stored?
output folder. Agentic Insight v2 persists structured artifacts to disk for context handoff and resumable runs.Thanks — we use this to improve the page.