# MS-Agent — Lightweight Agent Framework with MCP

> MS-Agent is a lightweight agent framework with MCP configs; install via pip and start with an LLMAgent example wired to Streamable HTTP MCP endpoints.

## Install

Save as a script file and run:

## Quick Use

1. Install:
   ```bash
   pip install ms-agent
   # or: pip install 'ms-agent[research]'
   ```
2. Set your ModelScope key:
   ```bash
   export MODELSCOPE_API_KEY={your_modelscope_api_key}
   ```
3. Run the MCP-wired example from the README (LLMAgent + `mcpServers`).

## Intro

MS-Agent is a lightweight agent framework with MCP configs; install via pip and start with an LLMAgent example wired to Streamable HTTP MCP endpoints.

- **Best for:** teams who want an agent runner that can consume MCP configs without building a custom harness
- **Works with:** Python; ModelScope API (per README); Streamable HTTP MCP endpoints
- **Setup time:** 15–35 minutes

## Practical Notes

- Quant: install is `pip install ms-agent` with an optional `[research]` extra per README.
- Quant: validate MCP wiring by calling one tool via an MCP endpoint and logging the round-trip latency (baseline for regressions).

## How to use it as a harness (not a demo)

The README example is a good start, but production usage needs guardrails:

1. **Pin MCP server configs** per environment (dev/staging/prod) and store them in version control.
2. **Separate credentials** from configs: env vars only, no secrets checked into repos.
3. **Replayability**: keep a small set of “golden tasks” you can re-run after upgrades to detect regressions.

If your goal is evaluation/security, wrap runs with:

- input/output logging,
- tool-call allow-lists,
- and a cost/latency budget per task.

### FAQ

**Q: Does it support MCP?**
A: Yes. The README includes an example `mcpServers` config wired into LLMAgent.

**Q: How do I keep runs reproducible?**
A: Pin configs + keep golden tasks + log tool calls and outputs.

**Q: What should I do first?**
A: Wire one simple MCP server endpoint and run a single task end-to-end before adding more tools.

## Source & Thanks

> Source: https://github.com/modelscope/ms-agent
> License: Apache-2.0
> GitHub stars: 4,244 · forks: 496

---

<!-- ZH -->

## 快速使用

1. 安装：
   ```bash
   pip install ms-agent
   # 或：pip install 'ms-agent[research]'
   ```
2. 配置 ModelScope Key：
   ```bash
   export MODELSCOPE_API_KEY={your_modelscope_api_key}
   ```
3. 按 README 示例运行（LLMAgent + `mcpServers` 配置）。

## 简介

MS-Agent 是轻量级 Agent 框架，可通过 MCP 配置接入工具；用 pip 安装后，按示例把 Streamable HTTP 的 MCP endpoint 写入配置即可跑通 LLMAgent。

- **适合谁：** 想直接消费 MCP 配置、又不想自建一套 agent harness 的团队
- **可搭配：** Python；ModelScope API（见 README）；Streamable HTTP 的 MCP endpoint
- **准备时间：** 15–35 分钟

## 实战建议

- 量化信息：README 给出 `pip install ms-agent`，并提供可选的 `[research]` extra。
- 量化信息：先用一个 MCP endpoint 调通一次工具，并记录往返延迟，作为后续回归基线。

## 把它当 harness 用，而不是一次性 demo

README 示例能跑通，但要上生产还需要护栏：

1. **按环境固化 MCP 配置**（dev/staging/prod）并纳入版本管理。
2. **凭证与配置分离**：只用环境变量注入，不把 secret 写进仓库。
3. **可复现**：维护一小组“金标准任务”，升级后重复跑，快速发现回归。

如果目标是评测/安全，建议额外加：

- 输入输出日志，
- tool-call allow-list，
- 每任务的成本/延迟预算。

### FAQ

**它支持 MCP 吗？**
答：支持。README 给出了 `mcpServers` 配置并接入 LLMAgent 的示例。

**如何保证可复现？**
答：配置固定 + 金标准任务 + 记录 tool calls 与输出。

**第一步做什么？**
答：先接一个简单 MCP endpoint 跑通单任务闭环，再逐步加工具面。

## 来源与感谢

> Source: https://github.com/modelscope/ms-agent
> License: Apache-2.0
> GitHub stars: 4,244 · forks: 496


---
Source: https://tokrepo.com/en/workflows/ms-agent-lightweight-agent-framework-with-mcp
Author: Agent Toolkit