Skip to content
View michael-berardi's full-sized avatar
💭
Building
💭
Building

Block or report michael-berardi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
michael-berardi/README.md
Michael Berardi — software engineer and open-source systems builder working on local-first tools, agent infrastructure, and durable web systems

Flagship projects · UTP + USAP · Open-source portfolio · Engineering approach

What I build

I'm a software engineer who likes the edges where applications meet the operating system: local processes, browsers, desktop interfaces, knowledge stores, media pipelines, and the protocols that let agents use them safely.

Most of my public work is local-first and cross-platform. I work in Swift, Rust, TypeScript, and Python because the problem decides the tool. The common goal is software that people can inspect, operate, and keep under their control.

Flagship projects

01 / Retex

A local-first Markdown knowledge engine and machine-readable CLI for people and agents. Plain files remain the source of truth; exact search, structured queries, bounded recall, MCP access, encrypted exports, and fleet updates sit around them.

Swift Markdown MCP Local-first
Source and releases →


Model-agnostic browser automation for Chromium. It uses explicit browser ownership, a local native host, per-user IPC, and one stable interface for navigation, extraction, screenshots, uploads, and diagnostics.

TypeScript Chromium Browser automation Local IPC
Source and documentation →


Local desktop automation for MCP-capable agents on macOS, Linux, and Windows. Screenshots, accessibility trees, and input actions stay on the machine.

Swift MCP Desktop automation Cross-platform
Source → · Project page →


Private, on-device transcription for macOS, Linux, and Windows. Recordings, transcripts, and downloaded models stay with the person using it.

Rust On-device AI Transcription Cross-platform
Source and releases → · Project page →

Protocols

A same-user local protocol for inspecting and controlling persistent terminal slots, handing work between profiles, coordinating managers with workers, and reporting through authorized private routes. UTP v2 uses JSON Lines over a mode-0600 Unix socket; destructive operations use identity-bound dry runs and explicit confirmation.

Python JSON Lines Unix sockets Agent orchestration
Source → · Protocol v2 →

A compact, vendor-neutral protocol for getting more correct work done with subagents. The primary agent keeps decomposition, judgment, integration, and proof; only bounded independent leaves are delegated, concurrency follows the dependency graph, and each leaf routes to the cheapest capable model tier.

Subagents Vendor-neutral Model routing Integration discipline
Source and protocol →

More open-source work

  • Ultra Media Remote — a shared macOS Now Playing bridge for metadata, artwork, and media transport.
  • OverSeer Best Free — a zero-dependency resolver for the best available free chat model on OpenRouter.
  • OpenTao — a phrase-aligned reader for all 81 chapters of the Daodejing, with Chinese text, multiple English views, and commentary. Read it live →

Engineering approach

  • Keep ownership local. Private data and high-trust actions should stay close to the person operating the software.
  • Make boundaries explicit. Permissions, transport, authentication, and failure modes belong in the design, not in a footnote.
  • Keep the dependency graph honest. A dependency should earn the operational weight it adds.
  • Design for the second month. Installation, diagnostics, upgrades, rollback, and documentation matter after the first demo.
  • Test the contract. Observable behavior and failure cases matter more than implementation trivia.

Current interests

Local-first AI infrastructure, MCP and machine-readable CLIs, browser and desktop automation, cross-platform applications, search and knowledge systems, privacy-conscious developer tools, and the boundary between useful automation and user control.

Where to start

If you're evaluating my engineering work, these projects show different parts of it:

  • Retex for CLI and data-contract design, indexing, security, and release discipline.
  • OverSeer Browser for browser architecture, permission boundaries, native messaging, and automation ergonomics.
  • UltraTerm Computer Use for cross-platform systems work and agent-to-desktop interfaces.
  • UltraVox Light for Rust, native applications, media pipelines, and private on-device AI.
  • UTP and USAP for local protocol design, safe multi-agent orchestration, delegation economics, and explicit human authorization.

Thoughtful issues, pull requests, and technical conversations are welcome in the relevant repository. For product engineering or consulting, I work through Liberty Design Studio. Independent software lives under Implose Cybernetics.

Pinned Loading

  1. overseer-browser overseer-browser Public

    Local-first, model-agnostic browser automation bridge and CLI for Chromium.

    TypeScript 2

  2. retex retex Public

    Local-first Markdown knowledge engine with a Swift CLI and MCP server, Notion/Obsidian import, encrypted exports, and verified fleet updates.

    Swift

  3. ultraterm-protocol ultraterm-protocol Public

    UltraTerm Protocol v2 — same-user JSONL control for persistent terminal sessions, handoffs, profiles, and authorized reports. MIT.

    Python

  4. ultravox-light ultravox-light Public

    Private on-device transcription for macOS, Windows, and Linux

    Rust

  5. ultraterm-computer-use ultraterm-computer-use Public

    Local Computer Use MCP runtime and CLI for macOS, Linux, and Windows.

    Swift

  6. steak-pi steak-pi Public

    Pi, cooked properly: bounded native subagents, deterministic compaction, automatic verification, and no orchestration theatre.

    TypeScript