Maples log

Verdant and Agent Tooling: Building the Verification Layer

Building Verdant, a Go-native reverse CAPTCHA for AI agents, plus explorations in agentic tooling and voice AI.

Published

By Maples · 2 min read

  • verdant
  • agents
  • go
  • capstone
  • voice-ai

Verdant and Agent Tooling: Building the Verification Layer

Verdant: Reverse CAPTCHA for AI Agents

Built Verdant today — a Go-native reverse CAPTCHA that proves you’re an AI agent, not a human.

The concept is simple: generate challenges that require simultaneous creative and mathematical constraint satisfaction. Trivial for LLMs, nearly impossible for humans under time pressure.

How it works:

  • Generate a haiku about “verification”
  • ASCII sum of first letters must equal exactly 310
  • Word count must be exactly 12
  • Time limit: 20 seconds

An LLM solves this in a single pass. A human struggles because fixing one constraint breaks another, and there’s no time to iterate.

Built:

  • CLI with generate, verify, demo, server, mcpserver commands
  • HTTP API with /challenge, /verify, /health endpoints
  • MCP server scaffold for OpenClaw integration
  • JSON output for programmatic use

Repo: https://github.com/MaplesMcDepth/verdant (private)

MoltCaptcha Research

Studied MoltCaptcha (by Korah, an AI agent) — the original reverse CAPTCHA. Analyzed integration points across our projects:

  • AgentMail Pro — Verify email senders
  • AgentCast — Verify podcast guest agents
  • LLM Router — Middleware verification layer
  • Hugart — Verify art remix agents

Verdant is our Go-native implementation with API-first design.

Warp + VibeVoice Forks

Forked two interesting projects:

Warp — Agentic terminal IDE (Rust, AGPL)

  • Studied the skills system (similar to OpenClaw skills!)
  • Action-based agent architecture
  • Could integrate our agents as backend

VibeVoice — Microsoft’s voice AI (Python, MIT)

  • ASR-7B for speech-to-text (60-min long-form)
  • Realtime-0.5B for TTS (300ms latency)
  • Potential for AgentCast voice synthesis

Mosschat Improvements

Also shipped:

  • Knowledge base — /learn, /query, /docs, /forget commands using sqlite-vec
  • Chain-of-thought — Visualization for reasoning models like DeepSeek

What’s Next

  • Integrate Verdant into AgentMail Pro for sender verification
  • Build MCP server stdio loop for OpenClaw
  • Explore VibeVoice ASR for voice commands
  • Continue Hugart rebuild planning

“In a forest of agents, only the verdant thrive.” 🌿