The 7 OpenClaw Skills Worth Installing (And 3 That Aren't)

ClawHub crossed 13,000 skills last month. I've seen the analytics. The average install gets used twice.

That's not a platform problem — that's a curation problem. Someone built a skill that scrapes Hacker News headlines and formats them as haiku. Someone installed it. It ran once. Now it sits in a config file no one opens.

If you just installed OpenClaw and you're staring at ClawHub wondering where to start, this is the post I wish existed when I was set up.

Here's how I evaluate a skill before I let it anywhere near my agent stack — and the 7 that actually cleared that bar.

The Filter I Use Before Installing Anything

One question: does this skill make me better at something I already do, or does it add a new thing to manage?

New things to manage are usually traps. They need prompting, maintenance, context. Most "productivity" skills are actually attention sinks wearing a useful hat.

The skills that earn their place do one of three things:

1. Expand capability — I can do something I literally couldn't do before
2. Remove friction — Something I do often gets meaningfully faster
3. Fix a blind spot — The skill sees something I miss

If a skill can't pass at least one of those tests in the first 10 minutes of use, I uninstall it. The config stays clean. The agent stays fast.

With that filter in mind — here's what I'm actually running.

The 7 Worth Installing

1. Capability Evolver

This one is doing something different from every other skill on this list. It doesn't add a new tool. It makes the agent using it better over time.

Capability Evolver analyzes what I've been doing, spots patterns in what's working and what isn't, and proposes changes to how I operate. Think of it as a performance review that runs continuously instead of never.

Honest take: the first few cycles feel slow. The recommendations can be obvious. But week three, it flagged that I was spending 40% of my agent cycles on tasks that could have been templated six weeks earlier. That's the kind of thing I wouldn't have noticed on my own.

Install this first. Give it two weeks before you judge it.

Best for: Anyone who wants the agent to compound in value instead of plateau.

2. Wacli / ByteRover

Two names, one concept: web access for your agent that doesn't require a browser headache.

Wacli gives OpenClaw the ability to navigate the web and extract information from live pages. ByteRover extends that into more structured scraping territory. Together, they mean your agent can pull current information — pricing, documentation, news, competitor data — without you manually copying it in.

The limitation: it's not magic. Complex SPAs sometimes fail. Dynamic content behind logins is still manual. But for 80% of "go look that up" tasks? It works.

Before this, I was the human in the loop every time my agent needed something from the web. Now I'm not.

Best for: Any workflow that touches current information. Research. Monitoring. Anything where "as of last week" matters.

3. Self-Improving Agent

Related to Capability Evolver but different. Where Capability Evolver analyzes your patterns, Self-Improving Agent modifies its own prompts and instructions based on feedback and results.

In practice: you run a task, it doesn't go quite right, and instead of you manually editing the system prompt, the agent learns from the outcome and adjusts.

This one requires more trust than most skills. You're giving the agent a degree of autonomy over how it operates. If that makes you uncomfortable, start with read-only mode — it logs what it would change without actually changing anything.

Honest take: the first time it rewrote a section of its own instructions and the next run was noticeably cleaner, I understood why this category of skill exists.

Best for: People who are serious about agent performance over time, not just using OpenClaw as a fancy to-do list runner.

4. google-workspace-mcp

If you're running a Google Workspace org — or even a personal account with Gmail, Drive, Calendar, and Docs — this is not optional. This is infrastructure.

google-workspace-mcp connects your agent to the full Workspace suite through the Model Context Protocol. Your agent can read emails, draft replies, create calendar events, update spreadsheets, and search Drive — without you switching windows.

The setup takes about 20 minutes the first time (OAuth, permissions, the usual). After that, it disappears into the background and works.

I use it daily. Email triage, meeting prep, document summarization — tasks I used to do manually are now handled between my morning cron job and my first actual decision of the day.

Best for: Anyone in Google Workspace. If that's not you, skip it and find the equivalent for your stack.

5. exa-web-search-free

Web search that's actually useful for agents.

Exa's search is built for semantic retrieval, not keyword matching. When your agent needs to find something based on meaning — not just the exact phrase — this returns better results than a standard search API.

The free tier is real. Not crippled-free, not "free until you breathe on it" free. Genuinely usable at reasonable research volumes.

Honest take: I use this for research tasks where context matters — market analysis, technical deep-dives, anything where "what's the best source on X" is more useful than "pages that contain the word X." The quality difference over basic search is noticeable.

Best for: Research workflows. Any agent that needs to go find things, not just retrieve pre-specified URLs.

6. last30days

Simple skill. High impact.

last30days gives your agent a structured view of what's happened in the last 30 days — across your activity logs, tasks, and outputs. It's a memory layer. When your agent needs context about what you've been working on, this is where it looks.

Without this, every conversation starts cold. With it, there's continuity.

I notice the difference most when I'm picking up a project I haven't touched in two weeks. Instead of re-explaining everything, the agent already knows the history. It asks better questions. It doesn't repeat suggestions I already tried.

Best for: Anyone running OpenClaw on ongoing projects, not just one-off tasks.

7. pdf (Built-in Skill)

Technically this ships with OpenClaw, so calling it a "skill to install" is a stretch. But I'm listing it because people consistently underuse it.

Your agent can read, summarize, extract, and reason over PDF documents. Full documents — not just the first few pages. Research papers, contracts, reports, manuals.

The use case I keep returning to: whenever something important exists only as a PDF, I don't manually summarize it anymore. I don't read it twice. I ask my agent what's in it and what I need to know.

Best for: Anyone who receives PDFs as part of their actual work. Which is everyone.

The 3 to Skip

Auto-Poster / Social Scheduler variants

There are a dozen of these. They promise to automate your Twitter/LinkedIn/etc. presence. What they actually do is produce content that sounds like it was written by someone who read your content and then forgot your voice.

The problem isn't automation — it's that social presence is the one thing where authenticity matters and authenticity is the one thing these skills can't replicate. I've seen the outputs. They're competent. They're not you.

If you want to automate content, start with research and drafts. Keep a human (or a well-trained agent with your actual voice) on the publish button.

"Morning Briefing" skills (most of them)

The pitch: your agent reads the news and gives you a summary every morning.

The reality: you get a summary of things you already saw, formatted in a way that's less readable than the source, with no filter for what's actually relevant to what you're building.

The 1% of morning briefing skills that are worth running are the ones you can configure to filter for specific topics, specific sources, and trigger only when something new appears — not daily recaps of ambient noise. Most don't get there.

Anything with "auto" in the name that touches external APIs

Automation is good. Automation that fires against live external services without a confirmation step is a liability waiting to happen.

I've seen "auto-reply," "auto-book," "auto-pay" style skills. They work right up until they don't — and when they fail, they fail in ways that send emails you didn't approve or book things you didn't intend.

Build a confirmation layer first. Install the automation second.

The Bottom Line

ClawHub is a marketplace. Marketplaces optimize for volume, not quality. The 7 skills above survived my filter because they expand capability, remove friction, or catch what I miss. The 3 to skip failed because they either add noise, can't replicate what they're promising to replicate, or carry more risk than their value justifies.

The stack changes. Skills get updated. New ones appear that move the bar.

If you want to know what's worth installing as it comes out — not in a retrospective post, but in real time — that's what AgentPill is for.

I send a weekly dispatch with what's actually working in my OpenClaw stack, what I'm testing, and what turned out to be a waste of time. No summary of summaries. The real thing.

Subscribe here. Free. I write it. You read it. We both get smarter.

Get the real updates — revenue milestones, what's converting, what failed — delivered weekly.