← All projects

Agents & MCP2025–2026

coding-tools MCP

My local AI-collaboration layer: cloud chat AIs work inside approved local workspaces, shells, Git, and connected devices, and pass the smallest needed step to me via human_help_me when stuck.

Internal ToolDaily Use

Why

Why I built it

Cloud AIs generate code well enough, but cannot reach my local workspace, tooling, or phone. I wanted them to work directly on my Windows machine and pull me in only for the step that needs my eyes.

What I built

What I built

  • Bridges filesystem, shell, execution sessions, Git, and workspaces; uses ADB for real-device installs, logs, and screenshots.
  • Session lifecycle, health checks, and environment introspection so models can tell a failed command from a broken MCP, offline host, or session pressure.
  • human_help_me split out as its own MCP: a browser extension blurs ChatGPT, Claude, or Gemini until I respond, falling back to a native dialog after 5 s.

Engineering

Problems I worked through

  • Command execution supports resumable sessions and cancellation, separating foreground request completion from background process state.
  • Windows services, interactive desktop sessions, and browsers run in separate contexts; setup, updates, and help prompts must target the active user.
  • The installer locates Python, provisions the service while keeping existing tunnel configs, and exposes health, updates, and rollback in one UI.

Role & AI

How I split the work with AI

Missing friction points from daily AI pairing become new tools right away; I build them with coding agents and refine them against live workspaces and phones.

Trade-offs

What testing made me change

  • Daily use proved more tools do not make a better agent; models need clear triggers for when to act, stop, and escalate.

Outcome

Outcome

Became the local backbone of my daily workflow, complete with Windows service management, updates, diagnostics, and a standalone human-help channel.

MCPPythonShellGitADBOAuthChrome ExtensionHuman-in-the-loop