Best AI Agent Skills for Social Listening (2026)

The paid social listening platforms cost four figures a month and mostly sell you a dashboard. What you usually want is narrower: what has been said about this topic in the last thirty days, across the handful of places that matter, with links.

That job has quietly become tractable from an agent. Not because listening got easier, but because a few open projects now do the boring half, which is authentication, date windowing and rate limits, and hand the model clean text.

The catch is that this category has a legal edge the rest of our directory does not. Half of these read platforms whose terms of service prohibit automated access. That is worth knowing before you install rather than after.

Two different jobs called the same thing

Most tools in this space claim all three. The ones below are honest about which they do, and I have grouped them accordingly.

Monitoring: open sources, no credentials

last30days

mvanhorn/last30days-skill, MIT, 58,887 stars, pushed 21 August 2026. The strongest entry on this page and the one to start with.

It researches a topic across Reddit, X, YouTube, Hacker News, arXiv, Techmeme, Digg and Trustpilot, all constrained to a rolling window. The window is the design decision that makes it useful. Ask a model about a topic and it blends 2023 and last week into one confident paragraph; this forces a date boundary and hands you sources inside it.

Several sources work with no API key at all. arXiv and Techmeme activate automatically once their small CLIs are on your PATH, and first-run setup installs those for you. Digg's story clusters arrive without X authentication the same way. Where auth is needed, it reads cookies from the Chromium family including Brave, Edge, Vivaldi, Opera and Arc, plus macOS Keychain and Linux pass.

Install, taken from the project's own README rather than our database:

/plugin marketplace add mvanhorn/last30days-skill

Then /plugin install last30days inside Claude Code. On other hosts, npx skills add mvanhorn/last30days-skill -g covers Codex, Cursor, Copilot and Gemini CLI. Use one method per machine; Claude Code does not deduplicate across install paths and you will get two /last30days entries.

Our database prints claude skill add mvanhorn/last30days-skill, which is not a command that exists. Use one of the two above.

Deep dive: /skills/last30days-skill

TrendRadar

sansan0/TrendRadar, GPL-3.0, 61,649 stars, Python, last pushed 17 July 2026. Multi-platform trend and public-opinion monitoring, aimed squarely at Chinese-language sources, with keyword-filtered push notifications to Slack, Bark, WeWork and others.

It runs three ways: GitHub Actions on a schedule, Docker, or locally. There are two published images, wantcat/trendradar for the monitor and wantcat/trendradar-mcp for the MCP service, the latter exposing analysis over HTTP on port 3333. Storage goes to local SQLite or anything S3-compatible, Cloudflare R2 included.

Two things to weigh. GPL-3.0 has consequences if you were planning to embed it. And the last push was July, which is fine for a scheduled monitor but means new sources are not arriving quickly.

Deep dive: /skills/trendradar

ToolsOfTheTrade

cjbarber/ToolsOfTheTrade, 17,148 stars, last touched May 2026. A hand-picked list of tools drawn from Hacker News, and not installable in any form. It is reading material, and our database's install string for it is meaningless.

I include it because it is a good manual counterpart to automated monitoring: when you want to know what practitioners adopted rather than what trended.

Deep dive: /skills/toolsofthetrade

Listening: the layer that needs your account

Agent Reach

Panniantong/Agent-Reach, MIT, 73,837 stars, Python, pushed 12 August 2026. Reads and searches social content so an agent can act on it. Documentation is Chinese-first, which is worth knowing before you commit to it.

The install behaviour is the reason it earns a place here over louder alternatives. agent-reach install defaults to checking your environment and listing what is missing; it will not install system packages or write configuration unless you pass --system. There is a --dry-run that previews every action, and an uninstall with --keep-config. For a tool that wants shell access and credentials, defaulting to read-only is the correct choice and a rare one.

It is a Python package: pip install agent-reach, then run the installer. Not an npx target, whatever our database says.

Deep dive: /skills/agent-reach

xiaohongshu-mcp

xpzouying/xiaohongshu-mcp, Apache-2.0, 15,417 stars, Go, pushed 20 August 2026. An MCP server for Xiaohongshu, the platform Western tooling ignores and which matters enormously if your market is Chinese consumers.

Deployment is Docker: docker pull xpzouying/xiaohongshu-mcp, or the supplied compose file. The maintainer points people who do not want to run containers at a lighter sibling project instead, and keeps a long troubleshooting issue open, which tells you the deployment is fiddly.

Deep dive: /skills/xiaohongshu-mcp

MediaCrawler

NanmiCoder/MediaCrawler, 63,401 stars, Python, pushed 14 August 2026. Crawls posts and comments across Xiaohongshu, Douyin, Kuaishou, Bilibili, Weibo and Tieba. Runs through uv, drives a real browser via Playwright, and ships an optional API server.

This is the most capable listening tool on the page and the one I would think hardest about. It is a scraper aimed at platforms whose terms prohibit automated collection, and GitHub does not resolve a standard licence for the repository. Personal research is one thing. Building a commercial product on it is a different conversation, and one to have with a lawyer rather than a directory.

Deep dive: /skills/mediacrawler

Where publishing tools sit

dreammis/social-auto-upload, 14,464 stars, pushed 19 August 2026, automates uploads to Douyin, Xiaohongshu, Video Accounts, TikTok, YouTube and Bilibili. It belongs to the third row of the table at the top of this page, not the first two, and it is here as a boundary marker.

Keep it on separate credentials from anything you use for reading. A listening tool that hits a rate limit costs you a retry. A publishing tool holding the same session, driven by the same agent, can post. The blast radius is not comparable, and the cheapest way to keep it that way is two accounts.

Deep dive: /skills/social-auto-upload

The generic scrape layer underneath

When no purpose-built connector exists, you fall back to fetching pages. Three entries cover most of it: Firecrawl for search-and-scrape with structured extraction, Scrapling for adaptive scraping that survives layout changes, and the Apify MCP server for renting someone else's maintained scrapers instead of writing your own.

Apify is the pragmatic choice for social specifically, because the maintenance burden of a social scraper is close to constant and paying for it is often cheaper than absorbing it. Note that the npm package there is @apify/actors-mcp-server, not the unscoped name our database prints.

Our agent skills for web scraping page goes deeper on that layer.

What I would run

• Start with last30days. It covers the open sources, needs almost no setup, and answers the question most people are asking.
• Add TrendRadar only if you need scheduled pushes or Chinese-language news coverage. It is a monitor, not a research assistant.
• Treat anything that reads a logged-in social account as a separate decision with its own risk. Use a secondary account, expect to be rate-limited, and read the platform's terms.
• Reach for Apify before writing a scraper. Your scraper will break in a fortnight; theirs is someone's job.
• Keep publishing tools out of the same credential. A read failure is an annoyance. A write failure posts something.

For the commercial platforms this replaces, see AI tools for competitive intelligence and AI tools for market research, plus the full list for social media managers.

FAQ

Is any of this legal?

Reading public news, forums and papers is uncontroversial. Automated collection from social platforms usually breaches their terms of service, and in some jurisdictions touches computer-misuse law once authentication is involved. The distinction that matters is not public versus private, it is whether you logged in and whether the terms you accepted prohibit it.

Which one should I install first?

last30days. It is the most maintained, the most useful without credentials, and the only entry here with an install path verified across five different agent hosts.

Why do these read my browser cookies?

Because most social platforms have closed or priced their APIs, so the practical route is reusing a session you already have. It works, and it also means the tool holds a credential that can act as you. Prefer a separate browser profile with a secondary account.

Do these replace Brandwatch or Meltwater?

For research questions, often yes. For anything needing historical archives, sentiment models you can defend to a client, or a contract that indemnifies you, no. The open tools answer questions; the platforms sell coverage and accountability.

Why are the star counts here so high for such niche tools?

Several of these are large Chinese-language projects with audiences our English-language directory rarely surfaces. The counts are real, checked against the GitHub API rather than copied from our own records. High stars here signal reach in a market you may not be in, not necessarily fit for your use case.