🪴 Brain

Open-Source AI Projects Roundup (Matthew Berman)

3 min read

Open-Source AI Projects Roundup (Matthew Berman)

🗒️ Description

Matthew Berman’s tour of four free GitHub projects you probably haven’t heard of — a trending-search skill, a local NotebookLM clone, an agentic-engineering skill, and a context-compression wrapper that can slash your AI bill. The recurring install pattern across all four: don’t touch the command line — paste the GitHub URL into your agent (Cursor / Codex / Claude Code) and say “install this.”

“I’ve basically stopped installing things through the command line. I simply copy-paste the URL directly into Cursor or Codex… and I say ‘install it.’”

🧩 The four projects

1. Last 30 Days — human-voted trending search

A Claude Code-style skill (install, then /last30days <query>) that works like an anti-Google. Instead of links + ads, it searches Reddit, Hacker News, Polymarket, GitHub, X, YouTube, and TikTok in parallel, scores results by what real people upvote/engage with, and an AI judge synthesizes one brief — “millions of people voting with their attention and their wallets.”

  • ~40k stars; by Matt Van Horn (co-founder of the company that became Lyft).
  • V3 engine resolves entities before searching — type openclaw and it maps to the right X handle and subreddits.
  • Can --emit=html a shareable brief. Good for recent, trending knowledge (e.g. researching a week-old term like “loop engineering”).

2. Open Notebook — local NotebookLM clone

A free, open-source, optionally fully local clone of Google’s NotebookLM: upload PDFs/links/docs, get grounded Q&A with citations, and generate a multi-host podcast from the content. Hosted (OpenAI) or local (OpenAI/Ollama/LM Studio); per-process model selection. → dedicated note: Open Notebook.

3. Agent-skills engineering workflow

A ~56k-star skill that gives you seven slash commands mapped to seven stages of engineering — spec → plan → build → test → review → code/simplify → ship — structured as a guided flow. Berman compares it to gstack (Garry Tan’s stack) but more narrowly focused on the engineering workflow rather than building a whole company. Starts with /interview-me to extract what you’re building into a markdown spec, then drives spec-driven development. (Distinct from this vault’s Agent Skills note, which covers Anthropic’s skill system.)

4. Headroom — context compression, up to 90% savings

Compresses everything the agent reads (tool outputs, logs, RAG chunks, files, history) before it reaches the LLM — same answers, a fraction of the tokens (47–92% in the demos). Wraps any harness (headroom wrap claude), reports savings (headroom perf), and mines failed sessions into CLAUDE.md/AGENTS.md (headroom learn). → dedicated note: Headroom. Caveats: installs Serena by default (--no-sa to skip), telemetry on by default.

✍️ Notable quotes

“You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agent.” — Peter Steinberger (surfaced via Last 30 Days)

“Same answers, fraction of the tokens.” — on Headroom

📒 Takeaways

  • Install-by-URL is the new norm — agents install their own tooling from a GitHub link; the CLI is optional.
  • Two of these target cost/quota directly — Headroom (context compression) is the standout for anyone hitting Claude Code limits; pairs with Caveman (output) and Token Optimization for Claude Code.
  • Recency is a gap LLMs have — Last 30 Days fills it with human-vote signal rather than model knowledge.
  • Sponsor: ElevenLabs (11 Labs) — for natural podcast voices in Open Notebook.

📖 Further reading/watching

Template: knowledge_note_info

Graph View

Backlinks