We built MCPtoAI, an open-source client for using MCP tools across AI models — looking for technical feedback
We’ve been building MCPtoAI, an open-source client for using MCP tools across different AI models without tying the workflow to a single provider.
The basic idea is:
- install the client on macOS / Windows, or use the Linux CLI
- add your own AI provider credentials locally
- connect MCP servers once
- configure individual MCP tools with Off / Ask / Allow permissions
- switch between supported models while keeping the same conversation and MCP setup
- use it from Desktop, the web, or a phone while tool execution remains controlled by the paired device
One of the main things we wanted to avoid was giving a remote AI session unrestricted access to local capabilities. Provider credentials, MCP OAuth tokens and permission settings stay on the device, and higher-impact actions still require explicit approval.
The client side is open source:
GitHub: https://github.com/mcptoai/MCPtoAI
Demo: https://youtu.be/nAhHaHcHs1U
Website: https://mcptoai.com/
We’re at the point where feedback from people who actively use MCP is more useful than continuing to polish things in isolation.
We’d especially like feedback on:
- whether the per-tool permission model makes sense
- how the local stdio vs remote MCP handling feels
- the device / relay architecture
- model switching while keeping the same MCP setup
- anything in the MCP workflow that feels unnecessarily complicated
This is obviously our own project, so this is self-promotion, but the main goal of posting here is to get technical criticism and find what’s missing before we push it more broadly.