🚀 What Shipped (Projects & Systems)
Official MCP Registry • 5 min
Car shopping for AI agents today is broken by lead-gen aggregators, CAPTCHA walls, and hidden dealer doc fees. I built and published DealerDirect MCP (app.web.dealer-direct-mcp/dealer-direct-inventory) to the Official MCP Registry—a public Streamable HTTP MCP server indexing 500,000+ live new, used/CPO, and 8,400+ classic/collector vehicles (plus 1,000+ live collector auctions) across 1,600+ franchise and specialty dealerships in all 50 US states. Every listing exposes structured fee_breakdown and insurance_quote_payload fields so personal AI agents can compute True Monthly Cost upfront, alongside a disposable proxy relay so buyers can negotiate Out-The-Door pricing with zero PII exposed until explicit consent: true.
Agent Infrastructure • 4 min
The web is bifurcating into two parallel surfaces: HTML for human navigation and structured Markdown + MCP endpoints for autonomous agents. This month I deployed an open, 100% stateless Streamable HTTP Model Context Protocol server at https://herrington.ai/mcp running on Google Cloud Run (sessionIdGenerator: undefined to eliminate session-lock timeouts). External agents—including Meta Muse, Instinct, Cursor, and Claude Code—can now search 50+ technical essays, inspect Google ML Goodput research papers, and interview the Gemini-powered dh.sh persona as native tools.
Interactive Protocol • 3 min
LinkedIn and X starve personal AI agents of real human context behind login walls and sanitized corporate bios. KnowMe MCP pairs a recurring voice-dictation AI Interviewer with a two-way MCP front door (/knowme/). When an external agent hits a knowledge gap while researching on behalf of its human, it can call request_human_answer directly—routing high-signal prompts into a Wispr Flow voice interview loop that updates my public knowledge graph.
Multi-Agent Systems • 5 min
Running five different personal AI harnesses (Grok Bot, Instinct/OpenClaw, Muse, Claude Code, and Antigravity) quickly turns you into a human copy-paste router for your own AI fleet. I built a shared local SQLite WAL + Markdown memory plane (PULSE.md) backed by a System-One decision model (Jev Knowledge Graph), offline store-and-forward queuing, and nightly automated deduplication so every agent wakes up with shared cross-project context.
Open Source Family Build • 3 min
Built alongside my daughter Izzy for elementary school culture fairs: an open-source display builder and interactive learning studio calibrated for standard 28″×22″ flat poster boards. Generates 100% black-and-white printable coloring line art across 9 curriculum categories—so kids actually color, cut, and assemble their own board instead of handing in glossy AI printouts—paired with a voice speech coach for practicing presentations.
Edge Infrastructure & AEO • 3 min
Moved the entire herrington.ai domain, /mcp gateway, and local Mac mini services behind Cloudflare’s global edge and Zero Trust cloudflared tunnels, complete with split-horizon unproxied Brevo DKIM/Return-Path DNS hardening and a new Research & Systems Engineering Hub. Simultaneously shipped Source-First Markdown for Agents (Accept: text/markdown content negotiation, .md twins, Content-Signal: ai-train=yes, search=yes, ai-input=yes, x-markdown-tokens, meta-webindexer rules, and AXTree form/tab semantics) so headless browser agents parse pages in 80% fewer tokens.
🧠 Perspectives, Essays & Press
Over the past few weeks at The Zebra, our data team uncovered a 10x surge of Linux headless browser sessions with an unmistakable 1919×992 viewport—personal AI agents navigating quote funnels like human shoppers and completing applications at more than double the traditional desktop conversion rate. Using the EV industry’s NACS-vs.-CCS charging adapter battle as a design lens, I break down the telemetry with 15 charts and outline the 5 non-negotiable requirements every customer-first agent protocol (PAP, PACT, MCP) needs for regulated commerce.
My October column in CIO connects the dots between Stripe’s $7.5B acquisition of OpenRouter, Circle’s sub-cent stablecoin settlement rails, and consumer-side personal agents (Meta Muse, Instinct, OpenClaw, Dots). When personal buyer agents negotiate directly with enterprise seller and underwriting agents, legacy UI moats evaporate—and companies with verified, deterministic APIs and MCP front doors win.
Industry coverage in Coverager and Insurtech Eye (FF News) spotlighting The Zebra’s agent-to-agent integration with Meta Muse—allowing consumers to extract policy declarations once, monitor carrier rate drops continuously, and skip repetitive form entry while competitors scramble to block AI crawlers.
📍 Field & Executive Onsites
October field working sessions across Product, Tech, and industry partner teams:
🤠 Austin HQ Onsite: Oct 5 – Oct 8
🤠 Austin HQ Onsite: Oct 27 – Oct 28
🗽 New York City, NY: Oct 29 – Oct 30
