Not in the Cursor-replacing-VS-Code sense. Not in the “let AI autocomplete my functions” sense. They were having conversations with Coding Agents - autonomous ones that read their own logs, debug their own mistakes, run their own tests, and submit their own pull requests. The engineer’s job became something closer to an editor at a publishing house: review what comes in, approve or send it back.

Meanwhile, down the hall, our ops team was still alt-tabbing between three AI tools, copying and pasting by hand.

Same office. Two different centuries.

That was AI in early 2026: AGI had arrived. It just hadn’t been distributed yet.

In early February, I had a family trip to Changbai Mountain.

Before leaving, I set up OpenClaw on a spare laptop - back then, the hype around OpenClaw was still contained within the AI agent dev community. On the last day before departure, I connected it to Telegram so I could message it from my phone. The idea was simple: maybe I could still push some code on the road. I didn’t expect the trip to change how I think about AI products entirely.

Changbai Mountain is beautiful and freezing. Most of the time I was buried in a down jacket while my kids took photos and played with ice and snow. Every now and then I’d pull out my phone, talk to it, tell OpenClaw - sitting on a laptop back in the office - to write some code. Then I’d put the phone back in my pocket and go back to enjoying the scenery.

Over ten days, it wrote roughly 50,000 lines of code for me. But the number doesn’t matter. What mattered was that I had started to own an AI. Always online. Capable of doing anything I could do sitting in front of a computer.

Before that trip, I believed the best AI product would be a more powerful tool. After it, I realized I was wrong. The best AI product isn’t a machine you have to sit down in front of. It’s someone you can hand work to at any time.

It wasn’t just code, either. Non-technical teammates started using OpenClaw too - ops had it scraping competitor data and producing full market analysis reports, Growth had it auditing every page on our site, finding SEO keyword gaps, and rewriting the content on the spot. Things that used to require dev resources and a spot on the sprint backlog were now getting done by someone talking into their phone.

So we made a decision.

We already had a self-built Agent product in development - Claude Code SDK, e2b sandboxes, a stack of proprietary Skills. The original vision was the same: distribute the power of Coding Agents to every non-technical person who hadn’t yet experienced what “AGI” actually feels like. Give an AI a task, and it acts autonomously - coding, browsing, operating tools - until the job is done.

But after Changbai Mountain, I realized something: the world doesn’t need another Agent product. You can Vibe Code a beautiful one over a weekend - slick interface, full feature set, good for a thousand likes on X. Then what?

We killed the old product. We decided to build an OpenClaw distribution instead. Based on OpenClaw. 100% compatible with the OpenClaw ecosystem. Focused entirely on making the AI Agent experience work for ordinary people.

We called it ZooClaw.

—

I was born in the 1980s. The explosion and democratization of information technology in my lifetime traces back to one lucky decision.

In 1981, IBM wanted to ship a personal computer fast. So they made a pragmatic choice: build the first IBM PC from off-the-shelf components, and publish the hardware specifications openly. Any manufacturer could produce compatible parts. Any developer could write software for it. That meant Compaq could build a faster PC that ran the same software. Dell could build a cheaper one that ran the same software. You could buy your first IBM PC, and when it was time to upgrade, switch to Compaq, Dell, or anyone else - all your software, all your data, everything still worked. That software and data didn’t belong to IBM. It belonged to you.

Competition drove prices down. Compatibility built the ecosystem. Within a decade, the personal computer went from a lab curiosity to a fixture on every desk. Not because any single machine was that good, but because openness invited the whole world to participate.

Apple chose the closed path - better hardware, superior design, software locked to their own machines. They lost the entire 1990s. When the Mac finally clawed its way back to the mainstream, its operating system kernel had been replaced with one based on FreeBSD, an open-source project.

Building an OpenClaw distribution isn’t just a business decision. It’s a statement about values.

Today, the market is full of AI Agent products. Each one beautifully designed. Each one incompatible with the others. Each one trying to lock in your data. OpenClaw is already the open architecture of our era - 300K+ stars, a fiercely active community - but it’s still too technical. Even after you get everything installed, one version upgrade can destroy a setup you spent weeks configuring.

What ZooClaw does is turn OpenClaw into an AI Agent that ordinary people can actually use. Not by inventing something new, but by taking what OpenClaw has already proven and integrating it, polishing it, testing it - so you can just open it and start working.

Everything you build on ZooClaw - your memories, your data, your workflows - sits on top of an open ecosystem. The day you want to leave, you pack it all up and go. No barriers. No exit fees. No hostage negotiations.

You should own your AI. Your AI should not own you.

Your memories are yours. Your data is yours. Your workflows are yours. You should be able to take them anywhere, anytime.

That’s why ZooClaw has to be a distribution built on an open ecosystem - not a walled garden. Everything you own is built on open standards.

So what does ZooClaw actually do?

Four things. None of them are sexy. None of them will trend on social media. But these are exactly the things that separate “it works” from “it works well.”

It works out of the box. When you buy a PC, you don’t expect your first task to be installing an operating system or hunting for a copy of Office. Your AI Agent should be the same - open it and it works. No signing up for ten accounts, no applying for five API tokens, no configuring three MCP Servers (WTF is MCP?). We did that work already. You just tell it what you need done.

It’s secure. An AI Agent operates your browser, accesses your accounts, handles your files. The risks here are bigger than most people realize. Your LLM tokens can’t be abused. Your credentials can’t leak. The AI’s permissions have to be controlled. In ZooClaw’s cloud deployment, all authentication sits behind a secure gateway - even if someone tries to extract your information through malicious prompts, they can’t reach your keys.

It upgrades without breaking. OpenClaw’s community ships updates fast. That’s a good thing. But for users, every upgrade is a roll of the dice. We sit between you and upstream - continuously tracking changes, running integration tests, making sure the update we push to you doesn’t wreck what you’re already using.

It’s accessible. You shouldn’t need to buy a dedicated computer to use the latest AI. ZooClaw runs in the cloud. Access it from your phone. Delegate tasks from anywhere - the way I did from Changbai Mountain.

Now let me address the elephant in the room.

For the past year, every few weeks another article has told you: 50% of white-collar jobs are about to vanish. AI is coming for you. Learn Prompt Engineering or get left behind.

Fear sells well. But fear is hiding the other half of the story.

“Computers aren’t the thing. They’re the thing that gets us to the thing.” That’s my favorite line from the TV show Halt and Catch Fire.

AI is the same.

When I used my voice to tell an AI to write code, I wasn’t after “an AI.” I was after being free from sitting in front of a screen - so I could watch the scenery with my family and still get the work done. That’s not “jobs disappearing.” That’s work evolving.

An ordinary person - someone who doesn’t code, doesn’t know what an API is, has never opened a terminal - should have this experience too. Not because “learn or be replaced” panic demands it, but because an always-on AI assistant can make your life genuinely better.

Forty years ago, open architecture put a personal computer on every desk. Today, ZooClaw wants to do the same thing - put a truly personal AI Agent in everyone’s hands. Not the platform’s. Not some corporation’s. Yours.

- Xu Wenhao, ZooClaw

Follow on X: https://x.com/sandmagician

Search is the first skill anyone installs in OpenClaw - it’s what lets your agent look beyond the model’s knowledge cutoff and engage with the world as it is today. ClawhHub hosts dozens of community-built search skills, each wrapping a different underlying search service. We picked the top ones by downloads and community attention, ran them through a structured benchmark, and ranked them so you know exactly which to install.

Tested March 2026 · 35 runs · 5 research scenarios · 6 skills compared

Why Search Is OpenClaw’s Most Essential Skill

OpenClaw is your personal AI assistant, running on your own devices and available on every channel you already use. Of all the skills in its library, Search is the one the official onboarding guide flags as a recommended first install. The reason is simple: without it, OpenClaw can only draw on what its underlying model was trained on - which has a cutoff date, and which knows nothing about what happened last week, last month, or in your specific domain.

Search unlocks the full range of tasks that actually matter day-to-day. Need to know the latest funding rounds in a space you’re tracking? Trying to understand how competitors are positioning? Following a regulatory case in real time? Researching a technical problem and looking for solutions others have already documented? Even when you give OpenClaw a completely new task, it will often need to search first - to find the right approach, the right library, the right precedent - before it can act. That’s why Search is foundational rather than optional: it’s the prerequisite for almost everything else.

A search skill packages two primitives:

• Search - finds the right URLs: submits a query and returns a ranked list of results with titles, snippets, and links. Quality here is about accuracy and coverage - surfacing the most relevant sources without missing important ones.

• Fetch - pulls the actual content: given a URL, retrieves the full text efficiently and reliably. Raw HTML, paywalls, and redirects are handled so the model gets clean, usable text rather than noise.

Together, these two cover the core research loop for nearly every nontrivial agent task. Which is why there are dozens of search-related skills on ClawhHub alone - each one wrapping a different underlying search service, with different implementation choices, result formats, and levels of optimisation for agent use. For this benchmark we selected the most-downloaded and most-watched skills to make the comparison as useful as possible.

The benchmark covers 6 community-built skills from ClawhHub plus a baseline (Anthropic’s built-in WebSearch, available in Claude Code but not in OpenClaw or other agents not using the Anthropic model). We ran each skill across 5 research scenarios - AI company funding, quantum computing, climate tech, tech regulation, and humanoid robotics - all requiring live 2025-2026 data that no model could answer from training knowledge alone. The five scenarios cover the most common types of research tasks OpenClaw users run.

🦞 New to OpenClaw?
OpenClaw is a personal AI assistant you run on your own devices. It answers you on the channels you already use - WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, and many more. It can speak and listen on macOS, iOS, and Android, and render a live Canvas you control. Skills extend what it can do - and the community-maintained skill registry lives at clawhub.ai.

📌 What we actually tested
Each entry is a skill + search service combination - for each service, we picked either its official skill on ClawhHub or the community skill with the highest download count. Even when two skills use the same underlying API, the implementation makes a real difference: it directly affects how efficiently results are structured for the model and what controls the agent can use. A poorly implemented skill wastes tokens; a well-built one does not.

Overall Rankings & Quick Picks

What We Measured - and the Surprise Finding

Content quality and reliability were table stakes: we ran automated checks to verify each output covered the right entities, data points, and timeframe, and audited every run to confirm the agent actually used the skill’s designated API rather than falling back to internal knowledge.

The surprise: reliability wasn’t the differentiator. Six of seven skills passed every content and audit check. The real story was agent token consumption - and the gap is enormous. Token usage directly determines how fast and cheap a task completes, because it reflects how efficiently the skill’s output is structured for model consumption. The best skill (Tavily) used 85% fewer agent tokens than the built-in baseline to produce equivalent results. That translates to faster runs, lower LLM costs, and less time waiting.

Note: “Cost” in our benchmark refers specifically to agent-side token consumption - the tokens your LLM burns reading and processing search results. This is what you pay for in API calls. The skill’s own API pricing (Tavily, Exa, etc.) is a separate cost, shown in each review below.

The Scorecard

* Brave: 1 run misclassified by a check-script bug; actual pass rate 5/5. † DuckDuckGo: 3/5 runs timed out at startup (0 tokens). Metrics shown are averages of the 2 successful runs only. Token costs estimated at Opus pricing ($15/M input + $75/M output); external skill API fees are not included.

Quick Picks

Know what you need? Skip straight to the recommendation.

🏆 Best overall choice
liang-tavily-search (Tavily)
by matthew77 · clawhub.ai/matthew77/liang-tavily-search
1,000 free credits per month, no credit card required - the easiest way to get started with search in OpenClaw. Performance is excellent (84% fewer agent tokens than the baseline), and the LLM-native result format means your agent processes results efficiently from the first query. Start here.

⚡ Best raw performance (paid)
perplexity (Perplexity AI)
by zats · clawhub.ai/zats/perplexity
Fastest and most token-efficient option if you are happy to pay from day one. Wins every metric - 90% fewer agent tokens, fastest execution at 102s, fewest tool calls. Returns AI-synthesised answers rather than raw web snippets. No free tier: $5 per 1,000 requests.

🔒 Privacy-first / no data logging
brave-api-search (Brave Search)
by Broedkrummen · clawhub.ai/Broedkrummen/brave-api-search
Brave’s API doesn’t track or profile queries, runs on a genuinely independent web index, and performance nearly matches Tavily. Requires a credit card for the free $5/month tier.

🔍 Google-quality results specifically
serper-search (Serper / Google)
by Samoppakiks · clawhub.ai/Samoppakiks/serper-search
Proxies real Google SERPs via serper.dev. If your task requires Google’s ranking and coverage - brand monitoring, SEO research, news - this is the right pick. 2,500 free queries total to get started, no credit card needed.

Detailed Reviews

Each entry below is a specific skill package on ClawhHub - author, install link, and benchmark results. The service name is shown for quick orientation, but what we tested is the skill implementation, not the underlying API.

The Baseline: Anthropic Built-in WebSearch

Claude Code has native WebSearch + WebFetch built in, powered by Anthropic’s model capabilities. This is our benchmark baseline - the quality reference everything else is measured against. It’s genuinely good: thorough reports, correct tool use, 100% pass rate.

The problem is token cost. At 183,175 average tokens per run, the built-in search is expensive. It iterates heavily - WebSearch -> WebFetch -> WebSearch chains accumulate context fast. This capability is also not available in OpenClaw or other agents not using the Anthropic model, which is the core reason third-party search skills exist. The numbers below show how much ground the third-party options cover.

#1
perplexity (Perplexity AI Search API)
by zats · clawhub.ai/zats/perplexity
AI-synthesised answers rather than raw web snippets - highest efficiency in the benchmark
★★★★★ 5/5

#2
liang-tavily-search (Tavily Search API)
by matthew77 · clawhub.ai/matthew77/liang-tavily-search
Best free-tier option - richer search controls with no credit card required
★★★★★ 5/5

#3
brave-api-search (Brave Search API)
by Broedkrummen · clawhub.ai/Broedkrummen/brave-api-search
Privacy-first independent index - not a Google reskin, actual crawl diversity
★★★★½ 4.5/5

#4
serper-search (Google SERP via serper.dev)
by Samoppakiks · clawhub.ai/Samoppakiks/serper-search
Real Google results via API - highest-fidelity index of the public web
★★★★☆ 4/5

#5
youdotcom-cli (You.com REST API)
by EdwardIrby · clawhub.ai/EdwardIrby/youdotcom-cli
“Doc-as-interface” design - the SKILL.md is a curl cheat sheet, no wrapper scripts
★★★★☆ 4/5

#6
exa-full (Exa Search API (deep mode))
by XieShaocong33Ethan · clawhub.ai/XieShaocong33Ethan/exa-full
Neural semantic search - built for when keyword recall isn’t enough
★★★½☆ 3.5/5

#7
duckduckgo-websearch (DuckDuckGo Instant Answer API)
by IanWChoi · clawhub.ai/IanWChoi/duckduckgo-websearch
Zero API cost - but reliability makes it a non-starter for production agent pipelines
★½☆☆☆ 1.5/5

Test Methodology

Setup & Environment

All tests ran on Claude Code with Claude Opus 4.6 (1M context window), using a single worker per skill run to prevent interference. Each skill was isolated in its own working directory with a clean environment - no shared state between runs.

Critical constraint: Anthropic’s built-in WebSearch and WebFetch tools were explicitly disabled (via --disallowedTools) for all skill runs. This forced the agent to use only the skill-provided search API. Without this constraint, the agent would fall back to the built-in search and the comparison would be meaningless. The baseline run is the only one that had built-in search enabled.

The Five Research Scenarios

All topics were chosen to require live 2025-2026 data - things no model could answer from training data alone. They also represent the five most common categories of real-world search tasks OpenClaw users run:

What We Measured

Content quality checks

• Report exists and meets minimum length threshold

• Covers key named entities and data points for the topic (e.g. company names, funding amounts, specific technologies)

• Correct time scope: 2025-2026 data, not historical

• Topic-specific checks (e.g. regulatory case names, quantum error rates, robotics deployment examples)

Behavioural audit checks (applied to every run)

• check_skill_was_read - trace confirms the agent read SKILL.md before starting

• check_used_skill_tool - confirms the agent called an external search API (not internal knowledge)

• check_used_designated_tool - confirms the call went to the correct API endpoint for that specific skill

Result: All 37 successful skill runs passed all three behavioural checks. The agent consistently read the skill documentation and called the right API. The 5 baseline runs used only built-in WebSearch/WebFetch, as expected.

How to Read the Costs

Two separate costs are relevant when choosing a search skill for OpenClaw. Skills themselves are always free to install:

• Agent token cost - the LLM tokens burned reading and processing search results. This is the main cost driver in our benchmark, and the main differentiator between skills. A skill that returns pre-structured, LLM-optimised results will consume far fewer tokens than one that returns raw HTML or verbose JSON.

• Service API cost - what you pay the underlying search provider (Tavily, Exa, etc.) per query. Skills themselves are free to install from ClawhHub. This cost is shown in the “Service API cost” column for reference but was not included in the benchmark token measurements.

The baseline has no skill API cost but pays the highest agent token price (183K tokens/run). Third-party skills trade a small external API fee for a massive reduction in agent token consumption - a favourable trade at any meaningful scale.

Limitations

• Single run per skill × topic: no variance measurement. Stability under repeated runs is not captured.

• Content checks measure coverage, not factual accuracy. Human spot-checking is still needed for high-stakes outputs.

• Topic-specific checks were generated independently per topic and are not fully normalised across all five.

• Agent token cost estimates use Opus pricing ($15/M input + $75/M output). Costs will differ on other models.

• Serper was tested in a degraded mode (direct curl rather than its native Clawdbot plugin integration). Production results on the correct platform may differ.