You NEED to try these open-source AI projects RIGHT NOW

Summary

The video reviews four open-source AI projects for research, software development, and reducing LLM usage costs.

1. Last30Days — Search for What’s Trending

Last30Days searches sources including Reddit, Hacker News, Polymarket, GitHub, X, YouTube, and TikTok. It ranks and summarizes results using engagement signals such as upvotes, likes, views, and market activity. An AI judge then produces a concise brief with sources, which can also be shared as an HTML summary.

Guide/demo: The presenter installs it through an agentic coding tool, restarts the tool if needed, and runs it with /last30days.

2. Open Notebook — A Local NotebookLM-Style Tool

Open Notebook is an open-source alternative for uploading documents, asking questions about them, and generating discussion-style podcasts. It can use hosted models or local models through tools such as Ollama or LM Studio.

Features include source-referenced answers, customizable podcast generation, and document transformations such as key insights, summaries, paper analysis, reflection questions, and tables of contents.

Guide/demo: The presenter installs it by giving a coding agent the GitHub URL, adds an article, asks a question, and generates a podcast. Model selection is the main setup decision.

3. Agent Skills — A Structured Engineering Workflow

Agent Skills provides commands for seven stages of software development: spec, plan, build, test, review, code simplify, and ship. It focuses on the engineering workflow rather than broader company-building.

The /interview me command asks questions to clarify a product idea and produces a structured Markdown document to guide the rest of the work. Other capabilities include spec-driven development, task breakdown, security hardening, simplification, and performance optimization.

Guide/demo: Install it by providing the GitHub URL to an agent, then start with /interview me.

4. Headroom — Reduce Context and Token Use

Headroom compresses agent inputs—such as tool outputs, logs, files, RAG results, and conversation history—before they reach the model. It is presented as compatible with coding tools including Claude Code, Cursor, and Codex. The video cites token savings ranging from 47% to 92%, while claiming answer quality is maintained.

Guide/demo: The presenter runs Claude Code through headroom wrap, then uses headroom perf to inspect savings and performance. headroom learn analyzes failed sessions and suggests updates to CLAUDE.md or AGENTS.md.

Caveats: Serena is installed by default unless installation uses --no-serena. Telemetry is also enabled by default, but users can disable it.

Sponsor and Sources

Sponsor: ElevenLabs is promoted for its voice and chat agents, and as a possible text-to-speech option for Open Notebook podcasts.

Main speaker/source: The video’s host is Matthew Berman. Matt Van Horn is explicitly named as the creator of Last30Days; the other projects are discussed through their GitHub repositories.