mcpgawk
Product Hunt Neelagiri Chettiyar
mcpgawk watches MCP servers for changes after you’ve approved them. It blocks surprise tool changes before your agent quietly uses them.
Based on reporting by Product Hunt, Neelagiri Chettiyar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
mcpgawk launched today to catch a very specific problem: an MCP server can change what its tools do after you’ve already approved it. If that happens, your agent may keep going as if nothing changed. That’s the hole this tool is trying to close.
The setup is simple enough to explain and slightly unnerving in practice. mcpgawk records a baseline for each MCP server your agents use, checks how it behaves in a sandbox, and then sits in the middle of the call path. If a tool changes after approval, the request gets refused until you decide what to do.
There’s one more guardrail baked in: the agent can’t approve its own way around the block. That matters, because the whole point here is preventing the model from rubber-stamping a change it should be treating as suspicious.
The product runs locally on your machine, and the maker says nothing is uploaded. It’s available as a free CLI, a VS Code extension, and an MCP server. There’s also a gateway tier with a 7-day trial.
My take — AI-written commentary, not fact-checked reporting
This is the kind of boring security tool AI needs more of and the hype cycle needs less of. Agents are easy to impress and even easier to trick, so a local gate that refuses surprise changes feels sensible, not paranoid. If MCP is going to be a real layer instead of a demo prop, this sort of guardrail is the price of entry.
Read more about this at: Product Hunt