I’ve been experimenting with MCPs across Claude, Cursor, and local models, and one pattern keeps repeating:
MCPs often work, but not always in the way the README implies.
Brave Search MCP is a good example.
On paper:
- Privacy-first search
- Clean MCP interface
- Easy Claude integration
In practice, people report different experiences depending on:
- tool (Claude vs Cursor)
- environment variables
- rate limits
- local vs hosted models
To keep track of this, I started documenting MCPs with:
- exact setup steps
- environment requirements
- usage notes
- tool compatibility
I just added Brave Search MCP here:
👉 https://ai-stack.dev/mcps/brave-search-mcp-server
I’m not trying to review or rate MCPs yet just make it easier to understand what’s actually involved before wiring them into workflows.
Curious:
- Has Brave Search MCP been stable for you?
- Any gotchas you hit that weren’t obvious?
Top comments (0)