经 AI Skill Hub 精选评估,FranxAgent工作流 获评「推荐使用」。这款AI工具在功能完整性、社区活跃度和易用性方面表现出色,AI 评分 6.8 分,适合有一定技术背景的用户使用。
FranxAgent工作流 是一款基于 Python 开发的开源工具,专注于 AI Agent、工作流编排、Python框架 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
FranxAgent工作流 是一款基于 Python 开发的开源工具,专注于 AI Agent、工作流编排、Python框架 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 方式一:pip 安装(推荐)
pip install franxagent
# 方式二:虚拟环境安装(推荐生产环境)
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install franxagent
# 方式三:从源码安装(获取最新功能)
git clone https://github.com/xhdlphzr/FranxAgent
cd FranxAgent
pip install -e .
# 验证安装
python -c "import franxagent; print('安装成功')"
# 命令行使用
franxagent --help
# 基本用法
franxagent input_file -o output_file
# Python 代码中调用
import franxagent
# 示例
result = franxagent.process("input")
print(result)
# franxagent 配置文件示例(config.yml) app: name: "franxagent" debug: false log_level: "INFO" # 运行时指定配置文件 franxagent --config config.yml # 或通过环境变量配置 export FRANXAGENT_API_KEY="your-key" export FRANXAGENT_OUTPUT_DIR="./output"
English | 中文
Let AI work for you like a partner – simple, secure, low cost. Now you can control the AI on your computer directly from your phone – no public IP, no port forwarding, one‑click secure remote access.
FranxAgent is a lightweight AI agent framework that enables AI to read files, execute commands, search the web, understand multimodal content, and truly interact with the world through the MCP protocol. v5.0.0 introduces a revolutionary Code Review Panel that transforms how you collaborate with AI on code changes – no more blind approvals, every edit is visible and editable before it touches your files.
---
| Tool | Purpose | Security / Notes |
|---|---|---|
time | Current date/time | Read‑only, safe |
read | Read file content or project structure | Read‑only. Code files return AST structure + line‑numbered content; directories return project skeleton. Supports documents, images, videos. |
write | Propose file changes (review‑before‑write) | **v5.0.0**: No longer writes files directly. Sends AI proposals to the Code Review Panel, where you review diffs, edit code, and approve changes. Supports overwrite, append, and edit modes. |
command | Execute system command | ❌ Direct deletion blocked; suggests moving instead. Supports timeout. |
search | Web search (DuckDuckGo) | Free, no API key. Returns title, snippet, URL. |
add_skill | Save a reusable skill | Saves Markdown skill file and immediately indexes it into the vector database. Zero restart, real‑time retrieval. No confirmation needed. |
MCP tool integration Add any MCP server (stdio mode) in config.json:
{
"mcp_servers": [
{
"name": "windows-mcp",
"command": "uvx",
"args": ["windows-mcp"]
}
]
} After startup, the AI automatically discovers all tools from these servers and calls them via the unified tools tool. No extra configuration – just say "take a screenshot".
Remote hardware (e.g., Raspberry Pi) To control remote hardware via SSH, add a configuration like:
{
"mcp_servers": [
{
"name": "raspberry-gpio",
"command": "ssh",
"args": ["-T", "pi@raspberry-ip", "python", "/home/pi/raspberry_mcp.py"]
}
]
}
Notes: - All tools (built‑in + MCP) are called via the single tools tool, saving tokens. - Built‑in tool names are fixed (e.g., read, write). MCP tools use server/tool format (e.g., windows-mcp/snapshot). - command has built‑in safety; deleting requires moving. - similarity helps with deduplication and checking. - Scheduled tasks run in background, daily repetition, no persistent user online required. - ett only supports Zhipu GLM models; ensure correct tools.ett configuration.
---
write tool no longer modifies files directly. It sends AI proposals to the frontend, where you review, edit, and approve changes – putting you firmly in control.---
read, write, command, search, add_skill, etc., extensible.knowledge/ are automatically merged into the system prompt, giving AI extra knowledge, rules, or workflows.command tool prohibits direct file deletion, suggests moving instead; high‑risk operations can require approval.config.json handles all settings.---
Windows users double‑click init.bat, macOS users double‑click init.sh – virtual environment and dependencies will be set up automatically.
Modify config.json according to your needs (see configuration section below).
In config.json, you can adjust the following parameters:
| Field | Type | Default | Description |
|---|---|---|---|
language | string | "en" | Language for UI and system prompt. |
api_key | string | - | API key (required). For Ollama, any value works. |
base_url | string | - | API base URL (required). For Ollama: http://localhost:11434/v1, for GLM: https://open.bigmodel.cn/api/paas/v4. |
model | string | - | Model name (required). Recommended: glm-4.7-flash, qwen2.5:7b, etc. |
settings | string | "You are a helpful AI assistant." | System prompt defining AI's role or behaviour. |
temperature | float | 0.8 | Randomness, range 0–2 (but recommended 0–1). Lower = more deterministic, higher = more creative. |
thinking | bool | false | Enable deep thinking mode (GLM models only). The model outputs reasoning steps but responds slightly slower. |
knowledge_k | int | 5 | Number of knowledge snippets retrieved per conversation for knowledge‑augmented prompts. Larger values inject more system prompt but may bring more relevant info. |
mcp_servers | list | [] | List of MCP server configurations, each with name, command, args (optional). Example: [{"name": "windows-mcp", "command": "uvx", "args": ["windows-mcp"]}]. |
Multimodal tool independent configuration (optional) Inside the tools field, you can specify separate parameters for ett (multimodal understanding). If not set, the top‑level configuration is used:
{
"tools": {
"ett": {
"api_key": "your-ett-api-key",
"base_url": "https://open.bigmodel.cn/api/paas/v4",
"model": "glm-4.6v-flash",
"temperature": 0.8,
"thinking": false,
"max_retries": 5
}
}
}
⚠️ Note: The multimodal toolettcurrently only supports GLM series models (e.g.glm-4.6v-flash). Ensure you have configured the correct API key and model name.
Example configuration (using GLM + Windows‑MCP):
{
"language": "en",
"api_key": "your-zhipu-api-key",
"base_url": "https://open.bigmodel.cn/api/paas/v4",
"model": "glm-4.7-flash",
"temperature": 0.8,
"thinking": false,
"knowledge_k": 5,
"settings": "You are a helpful AI assistant.",
"tools": {
"ett": {
"api_key": "your-zhipu-api-key",
"model": "glm-4.6v-flash",
"temperature": 0.8,
"thinking": false,
"max_retries": 20
}
},
"mcp_servers": [
{
"name": "windows-mcp",
"command": "uvx",
"args": ["windows-mcp"]
}
]
}
💡 Tip: After saving configuration changes, they take effect in the next conversation – no need to restart the service.
💡 Model recommendation: Useglm-4.7-flashfor conversation andglm-4.6v-flashfor vision tasks.
---
架构设计清晰,定位明确的轻量Agent框架。核心功能完整但项目仍在早期阶段,适合POC和学习用途,生产落地需补充文档和最佳实践。
该工具使用 AGPL-3.0 协议,商用场景请仔细阅读协议条款,必要时咨询法律意见。
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
⚠️ AGPL 3.0 — 最严格的 Copyleft,网络服务端使用也需开源,SaaS 使用受限。
AI Skill Hub 点评:FranxAgent工作流 的核心功能完整,质量良好。对于AI爱好者来说,这是一个值得纳入个人工具库的选择。建议先在非生产环境试用,再逐步推广。
| 原始名称 | FranxAgent |
| 原始描述 | 开源AI工作流:A simple and efficient AI Agent assistant。⭐6 · Python |
| Topics | AI Agent工作流编排Python框架智能助手开源 |
| GitHub | https://github.com/xhdlphzr/FranxAgent |
| License | AGPL-3.0 |
| 语言 | Python |
收录时间:2026-05-22 · 更新时间:2026-05-30 · License:AGPL-3.0 · AI Skill Hub 不对第三方内容的准确性作法律背书。