New data point that stopped me cold: one tool in this roundup slashed a code search from 17,000 tokens down to 1,400. That’s a 92% cut on a single call. Stack that across a full day of coding agents and you’re looking at real money saved, or hours more runtime before you hit your quota.
I found this while watching a breakdown from Matthew Berman, the creator behind the video, who dug up four free GitHub projects most people haven’t touched yet. He’s not the author of any of them. His value here is curation, and honestly he picked four winners. I was genuinely surprised how under the radar some of these still are. Let me break down what each one actually does and where you’d use it.
🔹 The four projects, quick version
- Last30days: a search skill that ranks results by human votes, not ad-driven algorithms. Built by Matt Van Horn, a co-founder of the company that became Lyft, it’s sitting above 40,000 stars.
- Open Notebook: a free, local clone of Google’s NotebookLM from the creator known as lfnovo, just under 30,000 stars.
- Agent Skills: a seven-stage engineering workflow skill from Addy Osmani, above 56,000 stars.
- Headroom: a context compressor from the developer chopratejas, around 24,000 stars and climbing fast.
Insight breakdown: why these matter
The thread connecting all four is that they fix problems the big hosted tools quietly create. Search engines bury you in ads. NotebookLM locks your documents on someone else’s servers. Coding agents burn through tokens like there’s no tomorrow. Each project here flips one of those defaults.
Take Last30days. Instead of crawling the whole web, it queries Reddit, Hacker News, GitHub, Polymarket, X, YouTube, and TikTok in parallel, then scores everything by upvotes, likes, and real-money betting odds. An AI judge synthesizes that into one brief. The original poster designed the V3 engine to figure out where to search before it searches. Type a niche term and it resolves the right handles and subreddits automatically. The result is recent, trending, human-backed info instead of SEO sludge.
Headroom works on a different layer. It compresses everything your agent reads before it reaches the model: tool outputs, logs, RAG chunks, files, and chat history. The mind behind it tested it on benchmarks like GSM8K, TruthfulQA, SQuAD V2, and BFCL, and quality held steady. Same answers, fraction of the tokens.
3 practical applications
- Research that’s actually current. Install Last30days as a skill, restart your agent, and run a slash command on any topic. You get a tight brief plus the source threads. There’s an HTML export too. Just add “emit=HTML” or ask in plain English for a shareable brief, and it spits out a clean page you can pass to a teammate.
- A private knowledge base with a podcast button. Open Notebook lets you feed in a PDF or a URL, then ask questions against it or generate a full multi-host podcast. The creator built in “transformations” too: extract key insights, build a dense summary, analyze a paper, or generate reflection questions. It runs on hosted models like the latest GPT, or fully local through Ollama or LM Studio.
- A structured build pipeline. Agent Skills maps seven slash commands to seven stages: spec, plan, build, test, review, simplify, ship. Start with the “interview me” command and it walks you through questions to pull out exactly what you want to build, then writes it into a clean markdown spec you reuse downstream. It’s narrower than a full company-builder toolkit, and that focus is the point.
Tips and pitfalls
- Easiest install path: skip the command line. The curator just pastes the GitHub URL into Cursor or Codex and says “install this skill.” Restart the agent afterward so new commands register.
- Open Notebook’s one fiddly step is picking which model powers which job. The video’s setup uses a top chat model, text-embedding-3-large for embeddings, and lighter models for transcription and transformations. Copy that and move on.
- Headroom ships with two annoyances the author flagged honestly. It installs a separate tool called Serena by default, which has nothing to do with compression. Pass the “no-sa” flag to skip it. Telemetry is also on by default, so disable it if you’d rather not share data. The code is open source, so you can strip anything you want.
- Headroom has two bonus features worth knowing. Run “headroom perf” to see per-model token savings and cache stats. Run “headroom learn” and it mines your failed sessions, then writes concrete corrections into your claude.md or agents.md file. In one run it found nine sessions across 378 calls and suggested an 8,000-token-per-session saving.
Quick pros and cons
- Pros: all free, all open source, most install in one paste, and the savings are measurable not theoretical.
- Cons: you manage your own model keys and config, and a couple ship with defaults you’ll want to turn off first.
What I like most is that none of these ask you to trust a black box. The data is human-voted, the documents stay local if you want, and the token math is right there in a performance report. That’s the kind of transparency I wish more paid tools offered.
If you run coding agents daily, Headroom alone could pay for your week in saved quota. Watch the full video for the live demos and exact setup steps, then grab whichever one fixes your biggest headache first.