Audience & trend research

last30days

An agent skill that searches Reddit, X, YouTube, Hacker News, Polymarket and GitHub at once, then writes one sourced brief

mvanhorn/last30days-skillPython60,541 as of 2026-08-31
By Jake Luo · Published Aug 31, 2026

last30days is an MIT-licensed agent skill that runs one question across Reddit, Hacker News, Polymarket, GitHub, X, YouTube and more at the same time, ranks what it finds by the engagement it earned rather than by search position, and has your agent synthesise a single cited brief. It installs into Claude Code, Codex, Cursor, OpenClaw and other skill hosts, and the free lanes run with no API key at all. For a founder it is a research shortcut: what people actually said about your market, a competitor or your own product in the last month, with the upvote counts still attached. It carries 60,541 GitHub stars as of 2026-08-31.

What it actually does

The premise is that the useful conversation about most topics is scattered across platforms that do not talk to each other, each with its own API, its own auth and its own idea of what a search is. Reddit holds the unfiltered opinion, Hacker News the developer argument, YouTube the long explanation, Polymarket a price on whether something will happen. No single search engine reaches all of them. last30days is the bridge: it queries the lanes you have enabled in parallel, scores results by the engagement real people gave them, merges duplicates across sources, and hands your agent a set of cited findings to write up. The output is a brief with thread scores and quotes in it, not a list of links.

LaneNeeds a key?What it adds
Reddit, Hacker News, Polymarket, GitHubNoWorks on the first run with no setup: threads with real upvote counts, front-page argument, market odds, and repository or contributor activity.
arXiv, Techmeme, DiggNo, but their small CLIs must be installedThe papers under the hype, the tech-press editorial layer, and clustered stories from a curated feed of high-signal accounts.
XYour own browser session or an xAI keyThe immediate reaction and the expert thread, separated into what a person posted and what was said about them.
TikTok, Instagram, Threads, PinterestA ScrapeCreators keyCreator and visual-culture signal with transcripts. Without the key these lanes stay dark and everything else still runs.

What we learned running it in production

We are not just users of this one. We vendor last30days into the container image that every AgentCeres customer's agents run inside, which means we operate it rather than merely install it, and two things there cost us real time. The first is the runtime: the engine wants Python 3.12 or newer and our base image ships 3.11, so we install a self-contained 3.12 alongside the system interpreter instead of upgrading in place — the document-building tooling that depends on 3.11 keeps working, and the skill's own preflight finds the newer interpreter by name on the path. Its one optional dependency, a segmenter for Chinese text, we install on a best-effort basis with a documented fallback, because a missing optional dependency should quietly degrade one lane rather than fail an image build.

The second lesson is conceptual and it cost us a broken scheduled job, so it is the one worth carrying away: this is a skill, not a tool. A tool is a function an agent can call. A skill is a set of instructions the model reads and then decides to follow. We wrote a recurring prompt that instructed an agent to call last30days the way it calls a connector, and the run simply failed, because there is nothing there to call. If you are wiring this into automation, ask for the research in plain language and let the agent reach for the skill itself. The same distinction explains why a skill is never free: everything installed competes for room in the agent's prompt, and a catalogue that outgrows that budget starts getting summarised before the agent ever sees it.

One more thing we can confirm independently, because we hit it from the other direction. The README describes rebuilding the free Reddit path after the public endpoint it used stopped working. We ran into the same wall on our own Reddit research and found the same answer: a general web-search tool with a last-30-days filter will happily return results that are not recent at all, and the reliable route to genuinely fresh threads is the platform's own recent-posts surface. If recency is the entire point of your research, verify that your dates are real rather than filtered.

Where it fits in growth work, and where it does not

Treat it as a listening pass, not a strategy. It answers what a community said and how strongly they reacted, which is a genuinely hard thing to get any other way, and it says nothing at all about whether those people will pay you.

  • Before you write. Angles grounded in an argument that actually happened, with the thread and its score attached, beat another summary of the top ten search results. Pair it with how do I come up with content ideas.
  • Before a comparison page. Reading what users of a competitor complain about in public is more honest raw material than reading that competitor's own marketing, and it dates itself, which their site does not. See how do I do competitor analysis for my startup.
  • Before you post in a community. Knowing what a subreddit has already argued about this month is most of what stops a first post reading like an advert. How do I get users from Reddit without getting banned covers the rest.
  • Not a replacement for customer conversations. Engagement measures what got attention, and attention selects for conflict and novelty. The quiet problem nobody posts about is often the one people pay to solve.
  • Not a keyword tool. A topic can be loud on social platforms and have almost no search volume, or the reverse. If the goal is ranking rather than reading the room, start at how do I do keyword research for my startup.

FAQ

What is last30days?
An open-source agent skill, MIT-licensed and written in Python, that researches a topic across Reddit, Hacker News, Polymarket, GitHub, X, YouTube and a growing set of other sources at once, scores what it finds by real engagement, and has your agent write one cited brief. It runs inside an agent host such as Claude Code, Codex, Cursor or OpenClaw rather than as a hosted product of its own.
Do I need API keys to use last30days?
No, not to start. Reddit, Hacker News, Polymarket and GitHub work on the first run with no key at all, and the free arXiv, Techmeme and Digg lanes only need their small command-line helpers installed. Keys and browser sessions unlock the rest — X wants your own session or an xAI key, and the TikTok, Instagram, Threads and Pinterest lanes want a ScrapeCreators key. Missing credentials skip a lane rather than break the run.
Is last30days a tool my agent can call?
No, and the distinction matters if you are automating with it. It is a skill: instructions plus scripts the model reads and chooses to follow, not a function exposed for calling. A scheduled prompt that orders an agent to invoke it like a connector will fail. Ask for the research in prose and let the agent decide to use it.
How is this different from asking a chatbot what people are saying?
Reach and evidence. A general assistant searches the web it can reach, which for most of these platforms is very little, and it answers from a blend of training data and whatever it found. This runs against each platform with your own access, keeps the engagement numbers, and cites the threads, so you can open the source and judge it yourself. The trade is setup and runtime: a real brief takes minutes and consumes a meaningful amount of model context.
Can I use it for competitor research?
That is one of its better uses, with a caveat. Comparing what people say about several products in one pass surfaces complaints and workarounds that neither vendor publishes, and the dates make it clear how current the picture is. The caveat is that unhappy users post more than happy ones, so read it as a list of things to verify with real customers rather than as a verdict.
Related projects
Marketing SkillsAn open-source collection of 45 marketing skills for Claude Code and AI agentsFirecrawlOpen-source web-data API — search, scrape, crawl, and map any site into clean Markdown or structured JSON that AI agents and apps can build onBrowser UseLet an AI agent drive a real browser — open pages, click, type, fill in formsClaude SEOOpen-source SEO audit plugin for Claude Code — parallel specialist agents that converge into one prioritized action plan

You built it. Now grow it.

AgentCeres is a managed AI marketing team — specialists draft the SEO, social, and outreach that fill your links, you approve what ships. 14-day free trial, from $39/month.

Start free trialMore projects