ScrapeX searches 10 social platforms plus the live web.
And it is one of four open-source AI agent research tools that now do on GitHub what the paid answer engines charge for: point an agent at Reddit, X, YouTube, GitHub and the open web. And bring back something grounded instead of guessed. If you want the short version: run ScrapeX for a self-hosted chat-style research agent, agent-search-tool for command-line access to 10+ platforms, Agent-Reach as a normalized access layer for building your own agent. And last30days-skill when you want one synthesized brief instead of forty tabs. All four live on GitHub, which means you can read every line before you trust a single result.
Why your agent needs threads, not just pages
The question you actually need answered as a builder is “does this work in production”. And that answer almost never lives on a landing page. It lives in a Reddit thread from someone three months into using the thing, a GitHub issue with a reproduction case, or a comment from a person who paid for the subscription and regrets it. Generic web search returns the polished version of reality.
These tools point your agent at the places where people vote with attention and money, which is exactly the framing the last30days-skill README uses: “That’s millions of people voting with their attention and their wallets every day.”
Here is my problem with that framing, and it applies to every tool in this list.
Engagement is a popularity signal, not a truth signal. last30days-skill describes itself as “An AI agent-led search engine scored by upvotes, likes. And real money – not editors,” which is genuinely great for answering “what do people actually think about this tool”. It answers nothing about “is it true”. Because a confident wrong take with more upvotes still outranks a correct one nobody read. These repos get you the raw signal cheaply. The verification stays your job, and that division of labor is the honest one.
The four tools, and what each one is for
ScrapeX: a self-hosted answer engine with receipts
ScrapeX describes itself as an open-source Firecrawl alternative: a self-hosted, chat-style research agent that extracts data from websites, social media, and APIs. The README pitches it as “think Perplexity. But it can see” Twitter/X, Reddit, YouTube, Bluesky, Hacker News, Mastodon, Instagram, TikTok, LinkedIn. And Facebook, all keyless through one API, alongside the live web. Answers come back with numbered citations you can click. Best for: replacing a generic answer engine with something you host and audit. The catch: a keyless claim covering ten platforms is a claim you test on your own machine before it touches anything you rely on.
agent-search-tool: honest, thin, inspectable
One install command gives an agent the ability to search and read 10+ platforms: Twitter/X, Reddit, YouTube, GitHub, Bilibili, LinkedIn, web pages, and more.
The internals are refreshingly unglamorous, which is a compliment.
Reddit is `rdt search “query”`, YouTube is `yt-dlp –dump-json URL`, GitHub is `gh repo view owner/repo`. And web reading is `curl https://r.jina.ai/URL`. If you want a paid search backend, Exa plugs in with `agent-search configure exa-key YOUR_KEY`. Best for: bolting cross-platform retrieval onto an agent you already run. The catch: every wrapper is one upstream change away from breaking. But you can read the exact line that broke.
Agent-Reach: the access layer, done right
Agent-Reach calls itself “a small open-source access layer for AI agents”: one consistent interface over GitHub, Reddit, YouTube, X, Instagram, and future connectors.
It normalizes search results into predictable objects and ships a runnable research-agent example you can read end to end.
Best for: building your own research agent on a clean foundation instead of writing a bespoke parser per platform. The catch is also the value, and it is the next section.
last30days-skill: one brief, not forty tabs
This one is an agent skill rather than a service. It researches a topic across Reddit, X, YouTube, Hacker News, Polymarket. And the web, searches those sources in parallel, then, in the README’s words, “an AI agent judge synthesizes it into one brief.” Best for: a repeatable weekly brief on your category. The catch: a single synthesized summary is exactly where errors hide. So read the citations before you read the prose.
| Tool | Best for | Sources | Watch out for |
|—|—|—|—|
| ScrapeX | Self-hosted chat research | Live web + 10 social platforms | Test the keyless claims first |
| agent-search-tool | Command-line retrieval | 10+ platforms, incl. GitHub, Bilibili | Wrappers break with upstream changes |
| Agent-Reach | Building your own agent | GitHub, Reddit, YouTube, X, Instagram | Recent X search needs a bearer token |
| last30days-skill | One synthesized brief | Reddit, X, YouTube, Hacker News, Polymarket, web | Errors hide inside summaries |
The auth table tells you what will break
Three lines from Agent-Reach’s documentation do more for your planning than any feature list.
Public Reddit search works without a token.
The GitHub REST API serves public data with an optional token, recommended. Recent X search requires `X_BEARER_TOKEN` and API access. That is the entire market in three sentences: Reddit and GitHub are effectively open for reading. And X is the walled garden where the free lunch ends.
Now hold that against ScrapeX’s keyless pitch across ten platforms and you see the central tension in this space.
Some connectors sit on genuinely open surfaces, some are thin wrappers hoping nobody changes the HTML. And the only way to know which is which is to run the thing.
The first thing I read in any research-tool repo is the auth section, as that is where the author admits what the feature list left out.
The token column predicts your maintenance burden better than any roadmap: the platforms that charge for access are the ones that change the terms on a schedule.
What this means if you run a lean operation
For a small business, this class of tool changes two jobs immediately. Competitive intelligence stops being a monthly vendor deliverable and becomes a query you run when you need it: what are users in your category complaining about this week, in their own words, on the platforms where they actually complain.
Content planning gets grounded the same way, since instead of guessing what your audience asks, you point the agent at the threads where they ask it and cite them.
The economics matter as much as the capability. When you run the code yourself, no vendor meter sits between your agent and its sources; your real costs move to the model doing the synthesis and the maintenance you own when a connector breaks. Be precise about what self-hosting buys you, though. The orchestration runs on your machine, but every fetch still hits someone else’s servers. You get auditability and control, not invisibility, and pretending otherwise is how people get surprised later.
Concrete next steps:
– Clone ScrapeX and run its chat interface on a question where you already know the ground truth, then check how many citations actually resolve.
– Install agent-search-tool and confirm the four out-of-the-box commands work on your machine before you build anything on top of them.
– Read Agent-Reach’s runnable research-agent example end to end; it is the fastest way to learn the normalized-object pattern.
– Save last30days-skill for a recurring brief, and always read its sources before trusting its summary.
Run one before you trust one
The pattern across all four repos is bigger than any single tool: retrieval is becoming a commodity you assemble from open parts.
And the differentiator is the judgment layer you put on top.
The paid answer engines sell you both in one bundle. The open-source stack sells you neither and hands you the parts instead, which is the better deal if you are willing to own the plumbing.
Pick one repo from the links above and run it against your own product category this week. Read the raw results before the synthesized summary, and count the citations that resolve. If the tool holds up, you just built a research capability you fully control. If it falls over, you found that out for free, which is the entire argument for running open source in the first place.
