Last30Days v3 Explained: Fresh Social Research for AI Agents Without Fake Sources

Direct Answer

Last30Days is not a magic bypass for the entire internet. It is an open-source research layer that lets an AI agent combine recent evidence from sources that are normally searched separately: Reddit discussions, X posts, YouTube transcripts, TikTok and Instagram activity, GitHub work, Hacker News, prediction markets, papers, and current web results.

The important upgrade in Matt Van Horn's new interview with Andrew Warner is not the number of sources. It is verification. Matt shows a failure mode every agent builder should care about: an AI can claim it ran Last30Days while quietly improvising the answer. The current project now includes Doctor mode, per-source status, a permission preflight, and clearer setup paths so you can check whether the evidence pipeline actually worked.

Interview credit: Andrew Warner and The Next New Thing. Guest and project credit: Matt Van Horn, creator of Last30Days. You can follow Andrew on X or contact the show at hi@thenextnewthing.ai.

Source Note

The interview and transcript provide the demonstrations, creator commentary, and examples. The factual setup spine is the current Last30Days README, the runtime skill specification, and the latest public GitHub release.

Checked on 14 July 2026, the repository had about 52,000 stars, 4,500 forks, an MIT license, and v3.14.0 as its latest release. This is a fast-moving project. Commands, providers, free allowances, and source behavior can change, so the repository should outrank this article when they disagree.

Link Map

What Changed in Last30Days v3

The original version was easier to explain: combine Claude Code with X and Reddit access, limit the window to 30 days, and synthesize what people are saying. The current v3 repository is closer to a research engine with a skill interface.

• Pre-research planning: the agent resolves relevant people, handles, repositories, subreddits, channels, and comparison terms before searching.
• Parallel source collection: supported sources run together instead of becoming a sequence of manual searches.
• Engagement-aware ranking: upvotes, comments, likes, freshness, and relevance help rank the evidence.
• Cross-source clustering: the same story from Reddit, X, and YouTube can be merged instead of repeated three times.
• Comments as evidence: top comments can be captured alongside the parent post or video.
• Doctor and diagnosis: health checks distinguish a quiet source from a source that never worked.
• Saved and structured outputs: briefs can be written to local files, emitted as HTML or JSON, and accumulated into a searchable research library.
• Discovery and monitoring: the latest release adds a nominate, enrich, and confidence-floor discovery flow for finding topics before they peak.

This expansion makes the tool more capable, but also more operationally demanding. A source can be unavailable, credentials can expire, a helper CLI can fall off the agent's PATH, a platform can change its markup, or a provider can rate-limit the run. That is why source status matters as much as the final prose.

The $3,000 per Month Claim

"Unlock $3,000/month in free data" is a useful video headline, not a verified credit balance. The project can replace part of a manual research stack and reduce the need for several separate listening or scraping subscriptions. That can be valuable. It does not mean every user receives $3,000 of provider credits or saves that amount each month.

There is also a live pricing detail worth checking. Parts of the Last30Days repository currently mention a large ScrapeCreators free-call allowance, while the ScrapeCreators pricing page showed 100 free credits, no card required, when this article was checked. Matt also mentions a different promotional value in the interview. Treat the provider's current signup and pricing page as the checkout truth.

Why Current Research Is Hard

Recent internet research is fragmented. Different models, search products, platforms, accounts, regions, and commercial agreements expose different slices of the web. A model may find a public page today and fail tomorrow. An old popular article may outrank the current answer. A platform search may expose posts but not comments. A transcript may exist but be blocked by rate limits or bot checks.

Last30Days addresses the fragmentation by making the user bring the access layer: free public endpoints where possible, local helper tools, optional provider keys, and optional authenticated browser sessions. It is a coordination layer, not a special legal exemption. Platform terms, access rules, rate limits, privacy obligations, and local law still apply.

The 30-day window also has a tradeoff. It is excellent for current reactions, tool changes, launch language, and emerging objections. It is weak for foundational facts, long-term performance, historical context, and questions where the best source is older than a month. Good research combines the fresh layer with primary documentation and durable evidence.

Source and Access Map

Doctor Mode and Fake-Result Checks

In the interview, Matt describes watching an agent say it had used Last30Days when it had not. The output looked plausible because a language model is good at writing the shape of a research brief. The fix is not a stronger prompt asking the model to be honest. The fix is observable execution.

The current skill includes three useful checks:

• Preflight: shows planned permissions, configuration, endpoints, optional commands, browser-cookie plan, and file writes before research. It does not read cookie values or run the research.
• Diagnose: reports the sources the runtime currently believes are available after resolving its configuration.
• Doctor: probes source health and returns per-source tiers with repair guidance for missing keys, expired sessions, missing helper tools, rate limits, or broken access.

From a source checkout, the repository documents commands in this shape. Installed paths vary by host, so asking the agent to run the bundled preflight or Doctor command is often safer than guessing its folder:

python3 skills/last30days/scripts/last30days.py --preflight
python3 skills/last30days/scripts/last30days.py --diagnose
python3 skills/last30days/scripts/last30days.py doctor

Install It in Claude Code, Codex, or OpenClaw

Use the current README before installing. These were the documented paths when checked:

Claude Code

/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30days

Codex and other Agent Skills hosts

npx skills add mvanhorn/last30days-skill -g -a codex

Remove -g if you want a project-scoped test instead of a user-wide install. For a first security review, project scope is the more conservative choice.

OpenClaw

clawhub install last30days-official

The interview also discusses running the skill inside Hermes and scheduled agent workflows. Compatibility is plausible because the project uses an Agent Skills structure, but install and permission behavior depends on the host. Confirm the current host documentation and run preflight before allowing network access or persistent schedules.

The Setup Tiers I Would Use

Why Comments Matter

Matt's strongest product point is that the comments can be more useful than the post. A launch video tells you what the creator wants to emphasize. The comments reveal confusion, skepticism, edge cases, alternative tools, and the exact language the audience repeats.

That does not make the top comment true. It makes it a high-value research lead. Use comments for voice-of-customer language, objections, humor, and questions. Use primary sources, reproducible tests, and official documentation for facts. For product work, the combination is powerful: official release notes tell you what shipped; comments tell you what users think it means.

The Workflows Worth Trying

1. Meeting preparation

Research a person, their company, and the conversation around both. Ask for recent launches, interviews, public objections, repositories, and three specific questions worth asking. Verify the identity resolution before using the brief. A common name can quietly produce a confident dossier about the wrong person.

2. Tool selection before building

Compare two or three tools inside the same window. Ask for official changes, recent user reports, known failures, migration pain, pricing changes, and the evidence behind the recommendation. The community can reveal operational friction that a feature table misses.

3. Prompt research after a model release

Find current examples from practitioners, then ask the agent to convert the findings into one ready-to-run prompt for your exact task. Do not blindly copy the most viral format. Test it against a baseline and keep the version that produces the best measurable result.

4. Current outreach research

Use recent discussion to understand a prospect's current priorities and the language buyers use around the problem. Do not use the tool to mass-harvest personal data or manufacture fake familiarity. The goal is a smaller, more relevant list and a message a human has reviewed.

5. Content and product discovery

Search for repeated questions, fast-growing topics, unresolved objections, and cross-platform disagreement. A strong idea has more than engagement: it has a real user, an expensive or frequent problem, evidence of attempted solutions, and a path to validation outside the feed.

Scheduled Agents and Monitoring

The repository supports stored runs, watchlists, daily or weekly briefings, and explicit Slack or webhook delivery. The interview also describes Last30Days inside Hermes and OpenClaw cron workflows. This is where the tool can become infrastructure rather than a one-off search.

Automation should compare deltas, not regenerate the same giant report forever. A useful weekly brief answers: what is new, what changed direction, which source first reported it, how confidence changed, and whether a human needs to act. It should also say when nothing crossed the evidence threshold.

• Run Doctor before the first scheduled research job and after repeated source failures.
• Set a maximum runtime, API budget, source list, and delivery destination.
• Keep raw evidence and source-status logs separate from the polished executive brief.
• Deduplicate repeated stories and alert only on meaningful changes.
• Require human approval before a finding triggers outreach, publishing, trading, or account actions.

Security and Evidence Boundaries

• Review the skill before installing. It runs code, makes network requests, can read optional browser-session data with consent, and writes research files.
• Use preflight first. The current spec says preflight reports planned permissions and writes without reading cookie values or running research.
• Prefer isolated credentials. Use dedicated API keys, spending caps, and a separate browser profile rather than exposing unrelated sessions.
• Know where files go. The current default is a Last30Days folder under the user's Documents directory unless an output path is configured.
• Separate signal from fact. Engagement, comments, and prediction odds measure attention or belief. They are not proof of an event or product claim.
• Respect platform and privacy rules. Public visibility does not remove terms-of-service, personal-data, copyright, or regional compliance obligations.
• Never hide partial coverage. A timed-out source is unknown, not negative evidence.
• Verify high-stakes claims independently. Medical, legal, financial, security, and reputational decisions need authoritative sources and qualified human review.

A 30-Minute First Test

1. Install Last30Days in a project-scoped test or another isolated environment.
2. Run preflight and record the planned network access, file writes, and optional credential paths.
3. Run Doctor. Save the list of active, degraded, and unavailable sources.
4. Choose one topic you already understand and ask for a seven-day window, not the full month.
5. Require five source URLs, publication dates, source status, and a separate list of facts versus community sentiment.
6. Open every cited URL. Check that the title, date, author, and claim match the brief.
7. Compare the answer with a normal web search and one official source.
8. Score useful findings, incorrect claims, missing sources, runtime, and API cost.
9. Only then add YouTube, X, or a paid social-data provider.

Bottom Line

Last30Days has grown from a clever Claude Code skill into a broad, open-source current-research layer for agents. The source breadth is useful, but Doctor mode is the more important feature. It turns a hidden collection process into something the operator can inspect.

Andrew Warner and Matt Van Horn show the practical upside: better meeting prep, current prompt research, tool comparisons, comment mining, and scheduled monitoring. The sober version is just as important. Source access is unstable, social signals are noisy, browser sessions are sensitive, and polished agent prose can disguise a failed run. Install narrowly, verify every source, keep primary documentation in the loop, and measure the tool by better decisions rather than the theoretical dollar value of its data.

Sources

• The Next New Thing: LLMs Can't Search TikTok or Reddit. This Free Open-Source Tool Does
• The Next New Thing on YouTube, official site, and Andrew Warner on X
• Matt Van Horn on X
• mvanhorn/last30days-skill
• Last30Days README
• Last30Days runtime skill specification
• Last30Days v3.14.0 release
• Agent Skills
• yt-dlp
• ScrapeCreators pricing and documentation
• mvanhorn/printing-press-library