Michael May

AI Engineer

Builder at heart. I design and ship AI systems: agents, LLM tooling, and the infrastructure that makes them reliable in production.

WHAT I'VE SHIPPED IN 2026...

Lately, I've been obsessed with building AI systems that actually ship.

A memory server I code with every day, an agent that books contractors over real phone calls, and a CLI that runs a swarm of coding agents at once. All shipped, all in use.

Jul 2026

Lexie

Shipped in under a week

An AI agent that makes the contractor calls you don't want to make. Describe the problem in a chat thread and Lexie fans out real outbound calls to local vendors in parallel, quotes the same scope to each one, negotiates the appointment, and books it. Every call is recorded and transcribed, and nothing dials until you confirm.

View project

Parallelism

I'm obsessed with using AI to build sharper AI tools. The pattern I keep returning to: give an agent a sandbox with exactly the tools and permissions its task needs and nothing beyond that, then compose those sandboxes into pipelines that run in parallel. Scope it right and a workflow that used to need a human in the loop runs on its own. It's the throughline in everything above.

Jun 2026

Parallely

Open source · MIT247 GitHub stars

Parallely earned 240+ GitHub stars from the developer community. It's a CLI I built that splits one large coding task into sections and runs them concurrently, one AI coding agent per section, all in a shared git worktree. A fullscreen TUI streams every agent's live progress, output, and token usage, and finished sections auto-commit to a shared integration branch. Backend-agnostic: it drives Codex, Claude Code, or OpenCode.

GitHub

Ongoing

MycelicMemory

Open source · MIT

I built MycelicMemory, an MCP server that gives any coding agent a robust, semantically indexed, knowledge-graph-backed memory system that is fully local and secure. Sub-100ms hybrid search, automatic capture from your coding sessions, zero cloud dependency, and it works with Claude, Cursor, Windsurf, or any MCP client. I always use it when I code.

View project GitHub

What do I use to code?

I use a mixture of several very powerful agentic tools made by the community, including but not limited to:

Backdoor★ 301

amux control-plane dashboard showing parallel agent sessions on a board

amux★ 321

berabuddies/agentflow repository on GitHub

AgentFlow★ 1.4k

The one I lean on hardest is amux, an open-source control plane I run as a modified fork, with Claude Code as its base. It supervises dozens of parallel Claude Code and Codex sessions, so I can kick off a run, watch every agent, and merge from my phone. Every session writes to MycelicMemory, my own MCP server for local, persistent memory across runs. Backdoor points Claude Code at any provider, I usually route through OpenRouter, and AgentFlow wires agents into a graph when a job needs real fan-out.

For browser work my go-to is agent-browser (Vercel Labs), the base for all my headless browser agents. Its plugin system lets them work through real user-auth flows and, where a site's terms allow, captchas.

Resume