DataFlow-WebUI is a full-stack open-source web application that provides a graphical interface for the DataFlow framework. It supports multiple AI code agents (Claude Code, Codex, Cursor) to help you build data pipelines through natural language.
If you already have Claude Code, Codex, or Cursor installed, just open this project and tell your agent:
请阅读 AGENT_SETUP.md 并帮我完成所有配置
The agent will run ./scripts/setup_all.sh, configure MCP, and get everything
ready. You only need Python 3.10+ and Node.js 20+ pre-installed.
git clone https://github.com/OpenDCAI/DataFlow-WebUI.git
cd DataFlow-WebUI
./scripts/setup_all.sh
./scripts/start.shsetup_all.sh handles everything: checks prerequisites, installs DataFlow +
backend deps + frontend build, and configures agent MCP connections. It's
idempotent — safe to re-run.
- Python 3.10+ with pip
- Node.js 20+ with npm (recommend nvm)
- Git
- At least one code agent CLI:
- Claude Code:
curl -fsSL https://claude.ai/code/install.sh | sh - Codex:
npm i -g @openai/codex - Cursor: Download Cursor IDE
- Claude Code:
DataFlow-WebUI supports multiple code agents in two usage modes:
These agents are spawned headlessly by the WebUI backend when you chat in the browser. Pick one from the dropdown next to the chat title.
| Agent | CLI binary | MCP config location | Auth |
|---|---|---|---|
| Claude Code | claude |
--mcp-config .mcp.json (passed by backend) |
ANTHROPIC_API_KEY (or ANTHROPIC_BASE_URL for 中转/gateway) |
| Codex | codex |
~/.codex/config.toml [mcp_servers.dataflow] |
OPENAI_API_KEY + optional OPENAI_BASE_URL，或 codex login OAuth（ChatGPT Plus，无需 API key） |
These agents run in your own IDE or terminal and connect to the WebUI's MCP server to push pipelines onto the canvas.
| Agent | How to use | MCP config location | Auth |
|---|---|---|---|
| Cursor IDE | Open this project in Cursor; the Agent panel auto-discovers MCP | .cursor/mcp.json (project-scoped) |
Cursor 内置认证 |
| Claude Code (terminal) | Run claude in this repo; it reads .mcp.json automatically |
.mcp.json at repo root |
ANTHROPIC_API_KEY |
Cursor 不通过 WebUI 前端调度。 它的使用方式是在 Cursor IDE 中打开本项目， Agent 会自动发现 MCP server，可以直接创建/执行 pipeline 并同步到 WebUI 画布。
# Claude Code
export ANTHROPIC_API_KEY=sk-ant-...
# 如使用中转:
export ANTHROPIC_BASE_URL=https://your-gateway/v1
# Codex — 方式一: API key
export OPENAI_API_KEY=sk-...
# 如使用中转:
export OPENAI_BASE_URL=https://your-gateway/v1
# Codex — 方式二: ChatGPT Plus OAuth (无需 API key)
codex login# If you used setup_all.sh, just:
./scripts/start.sh
# Or manually:
cd backend && uvicorn app.main:app --reload --port 8000 --host 0.0.0.0Once the backend is running, open your browser and visit:
http://localhost:8000/
💡 If you changed the backend port, replace
8000with your custom port.
In the chat panel header you'll see a small agent dropdown — pick Claude Code or Codex per session. Switching the dropdown starts a fresh conversation (different agents do not share session ids). Your last choice is remembered across reloads.
💡 如果你使用 Cursor IDE，无需在 WebUI 中选择 agent — 直接在 Cursor 的 Agent 面板中对话即可，pipeline 会通过 MCP 同步到 WebUI 画布。
| Symptom | Most likely cause | Fix |
|---|---|---|
<cli>: command not found in backend logs |
CLI binary is not on the same PATH the backend sees | Re-install or set DATAFLOW_<KIND>_CLI=/abs/path/to/cli before launching the backend |
Chat replies are empty / immediate done |
Agent failed to authenticate | Confirm the relevant API key is exported in the shell that started the backend; for Codex without API key, run codex login first |
Tool calls fail with MCP server not reachable |
The agent's MCP config doesn't point at this backend | Re-run ./scripts/setup_agent.sh <kind> |
| Agent invents non-existent operators | Construction skill not loaded | For Cursor, re-run ./scripts/setup_agent.sh cursor to regenerate .cursor/rules/; for Codex, regenerate AGENTS.md |
lang="zh" produces 0 records on English data |
Construction skill missing or stale | Same fix as above — the skill encodes the language-detection policy |
| Cursor IDE 中看不到 MCP 工具 | .cursor/mcp.json 缺失或 WebUI 后端未启动 |
运行 ./scripts/setup_agent.sh cursor 并确保后端在 localhost:8000 运行 |
You should now have the full DataFlow-WebUI stack running locally. If you encounter issues, double-check your Node.js and Python versions, and feel free to open an issue on GitHub.
Happy hacking! 🚀
If you use DataFlow in your research, feel free to give us a cite.
@article{liang2025dataflow,
title={DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI},
author={Liang, Hao and Ma, Xiaochen and Liu, Zhou and Wong, Zhen Hao and Zhao, Zhengyang and Meng, Zimo and He, Runming and Shen, Chengyu and Cai, Qifeng and Han, Zhaoyang and others},
journal={arXiv preprint arXiv:2512.16676},
year={2025}
}