BASHCAT
$ cd ../portfolio
AI Automation·2026

AI Agent & MCP Toolchain Development

A set of MCP servers and agent tooling that let AI agents such as Claude Code drive dev boards, USB and Bluetooth analyzers and browsers — and be steered remotely from Discord

clientOpen source
durationOngoing, 2025–2026
categoryAI Automation
stack
MCPFastMCPPythonTypeScriptClaude CodeDiscord APIArduino CLIUbertoothCynthion

Background

What an AI agent can do depends on the tools it can reach. Drawing on both our embedded and AI work, we wrap hardware instruments and development workflows that normally need a human at the keyboard into MCP (Model Context Protocol) servers the agent can call directly. All of it is open source on GitHub.

Published Tools

Hardware and Dev Boards

  • arduino-cli-mcp: compile, upload and manage boards and libraries through Arduino CLI, so an AI can write firmware and flash it onto real hardware
  • cynthion-mcp: drives a Cynthion USB test instrument so an LLM can capture, decode and emulate USB devices
  • ubertooth-mcp: drives an Ubertooth One for 2.4 GHz Bluetooth research — BLE sniffing, Bluetooth Classic discovery, AFH mapping and spectrum scanning — built on FastMCP with async capture sessions so long captures never block the agent

Web and Data

  • cloakbrowser-mcp: exposes a stealth Chromium over MCP with text, HTML, screenshot and interaction tools for agent web retrieval

Agent Infrastructure

  • dsh-discord-bot: maps each DeepSeek Harness workspace to a Discord channel. Send the agent work, watch which tool it is calling, read back the full trace with token usage and cache hit rate, and steer or stop it mid-run. The bot only opens an outbound WebSocket, so the agent's machine needs no open port and works from behind NAT or a corporate firewall
  • DSH plugins: import skills and MCP configs from Claude Code, Codex, Gemini CLI and others into DeepSeek Harness automatically, and estimate per-session cost live
  • tw-pii-redact: redacts Taiwanese personal data — names, national ID numbers, phone numbers and addresses. It hooks into agents such as Claude Code and Codex CLI, masking tool results before they reach the LLM and guarding commits

Design Principles

  • Thin but robust wrappers: wrap the official CLIs and SDKs instead of re-implementing hardware protocols, and return full error output so the agent can correct itself
  • Async long-running jobs: captures and sniffing run as sessions the agent starts and then polls
  • Safe defaults: scoped permissions, audit logs and PII redaction keep agents governable in real environments

Good Fit For

  • Letting AI agents operate internal systems, instruments or databases directly
  • Building MCP servers for existing APIs or equipment
  • Adopting agents with permission control, auditing and personal-data protection

More work in AI Automation.