[A follow up post] : SUNBURST SolarWinds Breach: Crime Scene Forensics #SolarWinds #SUNBURST

Blog post here : Repo here :

Releasing sniff v1 today after weeks of dogfooding. Every feature built with AI assistance gets scrutinized by the tool that helped build it. Available via GitHub and Homebrew. If you've ever wondered "Is this good code or just.good-looking code?" this is for you. Give it a.

We're at a critical point where AI coding assistants are getting more convincing. We need better ways to distinguish genuine progress from elaborate deceptions. Sniff is a tiny guardrail to keep your flow states in check.

This created a deterministic cycle: Plan → Implement → Sniff analysis → Fix deceptions → Verify completion. No more relying on AI self-reflection (which can be deceptive itself). Objective metrics replaced diplomatic "it could be improved" feedback.

The meta-loop was fascinating: I used Claude Code and AmpCode to improve sniff itself, with sniff watching every change. The tool started catching its own deceptions. "Your new feature is just an untested regex wrapper," it would report. Brutal but necessary feedback.

So I built sniff—a tool to detect these "confidence deceptions" in real-time. It catches unimplemented functions hidden behind abstractions, error handling that handles nothing, and comments that overpromise. Suddenly I could maintain flow state while having objective quality.

After months of AI pair programming, I noticed a pattern: code that looked sophisticated but fell apart under scrutiny. Functions that claimed comprehensive error handling but ignored all errors. TODOs that led nowhere. I was drowning in AI-generated confidence without substance.

The Great AI Reality Check: Why More isn't Always Better

I have an invite for #cometbrowser . let me know if anyone would like to give it a spin . its pretty decent (for one to DM email will get it).

Their implementations are often shoddy and unusable . The shelf life of their SDKs is not beyond an conference announcement.

RT @terminaltrove: mcp‑probe is a TUI tool for debugging & testing Model Context Protocol (MCP) servers. You can execute MCP commands, val….

RT @conikeec: @simonw [Updated docs with many more features for automated probing]. 🌐 Try it: 🔧 MCP Probe: The Swi….

MCP Probe is live:. 📷 GitHub: 📷 Crates: 📷 Homebrew: conikeec/tap/mcp-probe. Built this because I needed it. Sharing because debugging tools should match your workflow, not fight it. Terminal-native MCP debugging starts now 📷.

Install options (because choice matters):. ```bash.# One-liner (macOS/Linux).curl -fsSL | bash.# Homebrew. brew install conikeec/tap/mcp-probe.# Cargo. cargo install mcp-cli.```.Or download binaries. Whatever works.

**4/8**. But wait, there's more! 📷. It's not just a debugger. It's a full SDK. ```rust.use mcp_core::client::McpClient;. // Build custom MCP automation.let tools = client.list_tools().await?;.// Script all the things.```. One codebase. Two use cases. Zero compromises.

**3/8**. Interactive tool testing that doesn't suck:. • Navigate with arrow keys (like a human).• Guided parameter forms (required vs optional).• Execute → instant results.• Multiple view modes (JSON, tree, formatted).• Press 'R' for deep response inspection. No more.

**2/8**. What makes debugging painful?. 📷 Context switching (IDE → browser → forms).📷 Slow search through hundreds of tools.📷 Copy-paste responses to analyze.📷 No automation/scripting capability. MCP Probe fixes all of this. Press `/` → search 300+ tools instantly 📷.

**📷 Thread 1/8**. MCP testing was painful. Browser inspector → forms → copy-paste → repeat. My terminal was sitting right there. Mocking me. So I built MCP Probe 📷. Terminal-native MCP debugger that doesn't fight your workflow. 📷 Pure Rust • 📷 Zero deps • 📷 Actually.

My opinion on Apple's "The illusion of Thinking" paper