What does this skill do, and when should you use it?
This skill connects an agent to the MS-Agent Capability Gateway, a unified interface to ms-agent's projects, components, and atomic tools, exposed as MCP tools. It provides a rich set of capabilities including deep research, document research, financial research, web search (arxiv/exa/serpapi), code generation, video generation, LSP code validation (TypeScript/Python/Java), concurrent-safe file editing, and agent delegation. All project-level operations support an async submit/check/get pattern, allowing the agent to continue handling other messages while long-running tasks execute in the background. Requires the ms-agent Python package and MCP server configuration.
Executes MCP tool calls: submits deep research tasks, polls progress, retrieves reports; submits document research tasks and fetches reports; submits financial research tasks and fetches reports; performs web searches with selectable engines; submits code generation tasks and retrieves results; submits video generation tasks and retrieves results; delegates tasks to sub-agents (sync or async); checks directories or incrementally updates and validates files via LSP; replaces file contents via content-match or line-range. All async operations follow a submit/poll/get pattern.
- A researcher needing an in-depth investigation on a complex topic can submit an async research task and continue other work while it runs.
- A data analyst analyzing uploaded documents or URLs wants a structured research report with charts.
- A financial analyst wants to aggregate quantitative financial data and news sentiment into a professional report for a stock.
- A developer wants to generate a complete software project from a natural-language requirement and validate code correctness via LSP.
- A content creator wants to generate a short video (~5 min) including script, storyboard, voiceover, and subtitles.
- An agent needs to delegate a complex multi-step task to another agent and optionally cancel it if no longer needed.
How do you install this skill?
- The skill grants broad access to filesystem and external services without permission controls or user confirmation. Review configuration to ensure least privilege before deployment.
- The skill claims code and financial research run in Docker sandboxes, but there is no evidence of enforcement. Verify that sandboxing is mandatory.
- The skill depends on multiple external APIs (e.g., LLM providers, search APIs) that may have limited availability in mainland China. Ensure required services are reachable from your network.
- No rollback or undo mechanism is provided. Exercise caution with file editing tools and maintain backups.
- Publisher is unverified; independently assess security based on repository contents.
- Shell / CLI
- Network access
- Local filesystem
- MCP Server
Python 3.10+pipms-agent
First install the ms-agent Python package via pip (pip install ms-agent). Then configure the MCP server in your agent host's config: nanobot uses config.json, OpenClaw uses openclaw.json, Hermes uses config.yaml. Follow the exact JSON/YAML examples in SKILL.md.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/modelscope/ms-agent.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/ms-agent-skills" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
After configuration, all tools are available as MCP tools with a prefix depending on your host (e.g., nanobot: mcp_ms-agent_web_search, OpenClaw: ms-agent_web_search, Hermes: mcp_ms-agent_web_search, Cursor/Claude Desktop: no prefix). For async tasks, submit to get a task_id, poll progress, then fetch the final result. Refer to the Quick Decision Guide in SKILL.md to choose the right tool.
What are this skill's strengths and limitations?
- Rich toolset (30 capabilities) covering research, code, video, and more in one gateway.
- All project-level capabilities support async patterns, so the agent isn't blocked and can handle other messages.
- Works with multiple agent hosts: nanobot, OpenClaw, Hermes, Cursor/Claude Desktop.
- Code validation via LSP supports TypeScript, Python, and Java.
- Provides concurrent-safe file editing tools.
- Requires separate installation of the ms-agent Python package, adding deployment overhead.
- MCP server configuration is mandatory and differs per host, with a need to set PYTHONPATH.
- Async tasks require manual polling; there's no automatic completion notification.
- The SKILL.md doesn't mention a test suite or specific model compatibility.
- Some advanced features (financial research, video generation) may require additional API keys (EXA, SERPAPI, T2I) and can take hours to complete.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| MS-Agent Capability Gateway this page | 37 · Not recommended | ★ 4.4k | 20d ago | Apache-2.0 |
| Terminal Command Assistant | 51 · Use with care | ★ 10k | 4d ago | MIT |
| Vexor Semantic Code Search | 60 · Recommended | ★ 243 | 9d ago | MIT |
| LeanCTX Local Context Control | 59 · Recommended | ★ 3.9k | 3d ago | Apache-2.0 |
| NotebookLM Automation | 53 · Use with care | ★ 20k | 4d ago | MIT |
This skill is similar to Anthropic's Agent Skills implementation, but the main difference is that it exposes capabilities as MCP tools rather than filesystem-based skill directories.
How did FollowSkills review this skill?
The skill exposes a wide range of tools via MCP (deep research, code generation, video generation, file editing), but provides no permission controls, user confirmation mechanisms, or data-flow transparency. Documentation claims code generation runs in a Docker sandbox, but the enforcement and isolation are unverified. File editing tools can modify arbitrary paths without path restrictions or confirmation. Dependencies (e.g., akshare, pyright) are not security-audited. External effects are significant (e.g., generating videos, writing files), and no rollback mechanism is provided. Publisher is unverified, but source attribution is traceable (Apache-2.0, public repo). Score is below the anchor of 12, hence deductions.
SKILL.md and reference files provide detailed self-consistent instructions, including async patterns, parameters, and SOPs, but no executable or reproducible test evidence is provided. Although CI workflows and test files exist, the test file (e.g., test_partial_round.py) covers framework logic, not the skill's key paths (research, codegen, video). Thus, key path reproducibility is unverified. Error handling is documented, but actual failure cases are missing. Score is below the anchor of 10.
The skill clearly defines target audience (users needing deep research, code generation, etc.) and scenarios, and provides a quick decision guide for trigger conditions. However, non-fit boundaries are not explicitly stated, e.g., would it falsely trigger for simple summaries? Environment fit: dependencies on LLM APIs (e.g., ModelScope, OpenAI, Google) may have limited reachability from mainland China, and the docs do not address mainland China network access. Hence, deduct from the anchor of 15.
Documentation is well-structured with progressive disclosure (SKILL.md overview, reference files for details). Install and dependency notes are provided (e.g., pip install ms-agent), plus configuration examples for nanobot, OpenClaw, and Hermes. Parameters and tool naming are stable. Known limitations are disclosed (e.g., dependency on search API availability). License (Apache-2.0) and version (1.0.0) are given, but there is no changelog or maintenance responsibility statement. Publisher identity is unverified, but license and repo are public. Score is below the anchor of 15.
The skill describes capabilities (research, codegen, etc.), but provides no actual output examples or validation results. Claimed outputs like reports and videos are said to be directly usable, but there is no evidence they are verified. After completing the core task, outputs may require heavy rework, e.g., report quality depends on LLM. Thus, score is far below the anchor of 7.
Only author-provided documentation and examples, no third-party execution evidence. CI workflows and tests exist, but do not cover the skill's key paths. Claims cannot be independently reproduced. Score is far below the anchor of 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →