Multi-LLM-API-Gateway / README_MCP_HUB.md
Alibrown's picture
Upload 36 files
3060aa0 verified
|
raw
history blame
2.02 kB
# Universal MCP Hub (Sandboxed)
#### Universal MCP Server running in **paranoid mode** β€” built on [PyFundaments](PyFundaments.md) and licensed under ESOL.
The goal was simple: too many MCP servers out there with no sandboxing, hardcoded keys, and zero security thought. This one is different. No key = no tool = no crash. The Guardian (`main.py`) controls everything. `app/mcp.py` gets only what it needs, nothing more.
- MCP_HUB Built with Claude (Anthropic) as a typing tool. Architecture, security decisions
- Pyfundaments by Volkan Sah read [ESOL](ESOL)
---
## MCP Client Configuration (SSE)
To connect Claude Desktop or any MCP client to this hub:
```json
{
"mcpServers": {
"pyfundaments-hub": {
"url": "https://YOUR_USERNAME-universal-mcp-hub.hf.space/sse"
}
}
}
```
---
## Architecture
```
main.py ← Guardian: initializes all services, controls what app/ receives
└── app/mcp.py ← Sandbox: registers only tools with valid keys
β”œβ”€β”€ LLM tools (Anthropic, Gemini, OpenRouter, HuggingFace)
β”œβ”€β”€ Search tools (Brave, Tavily)
β”œβ”€β”€ DB tools (only if DATABASE_URL is set)
└── System tools (always active)
```
**The Guardian pattern:** `app/mcp.py` never reads `os.environ` directly.
It receives a `fundaments` dict from `main.py` β€” and only what `main.py` decides to give it.
---
## Security Notes
- All API keys loaded via Secrets (env vars) β€” never hardcoded
- `list_active_tools` returns key **names** only, never values
- DB tools are read-only by design (`SELECT` only, enforced at application level)
- Direct execution of `app/mcp.py` is blocked by design
- Built on PyFundaments β€” a security-first Python architecture for developers
> PyFundaments is not perfect. But it's more secure than most of what runs in production.
---
## License
Apache License 2.0 + [ESOL 1.1](https://github.com/VolkanSah/ESOL)
---
*"I use AI as a tool, not as a replacement for thinking."* β€” Volkan KΓΌcΓΌkbudak