YT-Proteus-bot/MineAgent

What it does

MineAgent is a local, frontend-backend split application for building Agent workflows on a browser canvas. You choose a framework, drag nodes, connect them, edit each node's Python code, and configure models; saving compiles the whole canvas into a standalone Python Agent project under projects/<canvas name>/. Running compiles and executes that generated project on your machine, streaming per-node events and model tokens back to the canvas (README).

The backend is FastAPI with SQLite; the frontend is React 18, TypeScript, React Flow and Vite (README). The canvas ships with 110 node types across categories such as RAG retrieval, tools, memory, flow control, multi-agent collaboration and framework-specific nodes (node list). Generated projects include agent.py, agent.json, runner.py, runtime_sdk.py, run scripts and skills/; agent.py is the framework-native implementation, and saving is incremental, preserving nodes whose code was manually edited (generated project layout, sync behavior).

Supported frameworks and modes

LangChain/LangGraph, AutoGen and CrewAI are listed as runnable; MetaGPT, Agno, SmolAgents, Semantic Kernel, OpenAI Agents SDK and Dify Workflow generate architecture diagrams and code skeletons but do not have native execution (framework table). Choosing a reasoning or collaboration mode, such as ReAct, Plan-and-Execute, supervisor dispatch or debate, generates a minimal working canvas for that mode rather than only setting a value (mode templates).

Quick start

On Windows, double-click start.bat or run start.ps1 from the repository root; optionally specify a port or development mode:

.\start.ps1
.\start.ps1 -Port 8080
.\start.ps1 -Dev

The script creates a Python virtual environment, installs backend and frontend dependencies, builds the frontend, starts the service and opens http://127.0.0.1:8000 (start instructions). On first launch with no canvases, a LangChain architecture diagram is created automatically.

If you have no API key, start the bundled mock model in a second terminal:

.\.venv\Scripts\python.exe tools\mock_llm.py --port 8012

Then set the Base URL to http://127.0.0.1:8012/v1, the model to mock-model and any value as the key under 用户中心 → 模型配置 to exercise the full canvas-to-code-to-run flow (mock instructions). For a generated project, run run.ps1 <input> for the zero-dependency preview or install the framework requirements and use run.ps1 -UseFramework to execute agent.py (standalone run).

Constraints

• Windows-oriented startup: the documented one-click entry points are start.bat/start.ps1; generated projects include both run.ps1 and run.sh.
• Framework execution depends on installed packages: if the selected framework's dependencies are absent, running falls back to the zero-dependency preview runner, which is labeled in the run panel rather than executing agent.py (run behavior).
• Only three frameworks are listed as runnable: the others produce diagrams and skeletons only (framework table).
• Node availability varies: document and multimodal nodes (pdf_parser, ocr_extract, speech_to_text, image_analyzer) rely on optional libraries and return an installation hint when missing (node list).
• Local data and keys: canvases, settings and run records live in local SQLite; API keys are encrypted in the local database and injected into subprocesses via environment variables rather than written into project files (model configuration).
• Community template uploads need GitHub setup: uploading a canvas as a template requires a public template repository and a GitHub OAuth App whose callback URL is http://127.0.0.1:8000/api/github/callback (template sharing).
• No license is declared in the repository metadata, so reuse and commercial terms are not established by the source.

HotRepo's take

MineAgent entered the Python new-project list on 2026-11-05 with 104 stars gained in roughly 9.7 hours since the recorded growth window began at 06:01 UTC, ending at 15:45 UTC; it has no recorded GitHub all-language daily rank. One snapshot is enough to note that early burst but not to describe a trend. What distinguishes the project in its own documentation is the scope of the generated artifact: it does not stop at a diagram, but emits a runnable project with framework-native agent.py, a zero-dependency preview runner, and explicit fallback behavior when framework dependencies are missing. The practical caveat matches that breadth: only three of the listed frameworks execute natively, so canvases built on the remaining frameworks should be treated as skeletons first.