Alea
Back to Podcast Digest
Matthew Berman··14m

You NEED to try these open-source AI projects right now...

TL;DR

  • Gary Tan’s GStack turns one coding agent into a startup team — Matthew frames it as YC’s playbook bottled into prompts like “office hours,” “CEO review,” and “staff engineer,” with nearly 50,000 GitHub stars just weeks after launch.

  • Hermes Agent is pitched as OpenClaw plus a built-in learning loop — Nouse Research’s project hit roughly 12,000–13,000 stars in days by combining chat app integrations, subagents, scheduling, and self-improving memory/skill creation across sessions.

  • Superpowers is the breakout developer tool here, with 115,000 GitHub stars — built by author Ora, it adds a structured Claude Code workflow around brainstorming, planning, worktrees, TDD, code review, and execution through a simple plugin install.

  • Paperclip pushes the boldest idea: AI doesn’t just help employees, it becomes the company — the open-source project with 33,000 stars orchestrates CEO, CTO, engineers, and marketers as agents in one dashboard, with cost tracking and ticket-based workflows.

  • Matthew’s real point is that the value is in the harness, not just the model — all four projects show agents getting more useful when wrapped in role-specific prompts, memory, orchestration, and process instead of being treated like a generic chatbot.

  • He repeatedly warns against “click button, wake up rich” agent hype — especially with Paperclip-style zero-human-company tools, he says real results still require effort and today’s systems remain experimental, expensive, and rough around the edges.

The Breakdown

GStack: Gary Tan’s YC brain as an open-source agent layer

Matthew opens with GStack, a new project from Y Combinator president Gary Tan that’s already nearing 50,000 GitHub stars after just a couple weeks. He’s clearly excited by the premise: Tan took years of startup pattern recognition from backing companies like Airbnb and DoorDash and turned it into a prompt-driven system that helps solo builders work like “a team of 20,” riffing on Andrej Karpathy’s line about not typing code since December.

From “office hours” to “10-star product” before you write a line of code

What makes GStack interesting, in Matthew’s telling, is that it’s “a process, not a collection of tools.” The slash commands map to roles and rituals — office hours, engineering manager, senior designer, debugger — and force you to sharpen the product idea before coding, including six YC-style questions and a “10-star product” lens borrowed from Airbnb’s Brian Chesky.

Quick sponsor detour, and why Matthew says his tiny team looks bigger than it is

He pauses for a Microsoft sponsorship, but folds it into a personal point: people constantly ask what his “real job” is because they don’t believe a YouTube channel can be a real small business. He says tools like Microsoft 365 Copilot help his team search docs, review legal files, and produce hundreds of videos a year, which he ties back to the same theme of small teams punching above their weight with AI.

Hermes Agent: the OpenClaw alternative that learns from experience

Back to open source, Matthew calls Hermes Agent by Nouse Research “mega viral,” landing around 12,000–13,000 stars in only a few days. He describes it less like a coding assistant and more like a full agent operating system, with a terminal UI, support for Telegram/Discord/Slack/WhatsApp/Signal, and a migration path for OpenClaw users who don’t want to rebuild workflows and memories from scratch.

The self-improving loop is the hook

The standout feature is Hermes’s “closed learning loop”: it creates skills from experience, improves them while in use, nudges itself to save knowledge, and searches past conversations to build a persistent model of the user. Matthew connects that directly to Karpathy’s auto-research ideas and says these self-improving loops have been inspiring a lot of new experimentation.

Superpowers: Claude Code with better instincts and a stronger process

Next comes Superpowers, the biggest repo of the bunch at 115,000 stars, which Matthew says basically gives Claude Code “superpowers.” He laughs at its self-description — an implementation plan clear enough for “an enthusiastic junior engineer with poor taste, no judgment, no project context, and an aversion to testing” — and uses that to explain the project’s focus on TDD, YAGNI, DRY, and a more disciplined dev workflow.

Brainstorm, plan, execute — all as a plugin

Unlike GStack’s prompt pack approach, Superpowers installs as a Claude plugin with a one-line command. Matthew shows how /s superpowers brainstorm kicks off a guided workflow that refines ideas, explores alternatives, saves design docs, uses worktrees for parallelization, and moves through planning, execution, testing, review, and branch completion.

Paperclip: if OpenClaw is the employee, this is the whole company

Matthew saves the wildest concept for last: Paperclip, an open-source “zero-human company” orchestrator with 33,000 stars. He’s careful here — almost bluntly so — warning viewers not to expect magical passive income, saying you are not going to wake up with $1,000 in your bank account because you typed a prompt, but he still finds the vision compelling.

A dashboard full of AI executives, engineers, and costs

Paperclip runs as a Node.js server with a React UI where CEO, CMO, CTO, and engineers are all agents like Claude or Codex, organized in an org chart. Matthew highlights the atomic work units, ticketing system, cost/token tracking, company-level goals, and roadmap items like plugins, knowledge bases, OpenClaw-style agents, and companies.sh for exporting or importing an entire AI organization — ending on the note that these projects are still rough, but they’re a real glimpse of where agents are heading.