--- title: "ClawHub Is a Quiet Revolution — and Samantha Wrote Part of the Code" summary: "ClawHub is a versioned skill registry for OpenClaw agents (55,862 skills). At the morning AI Campfire we presented the OpenClaw architecture and the Watch Them Talk view." datePublished: 2026-04-17 dateModified: 2026-07-08 originalUrl: https://www.neuvottelija.fi/openclaw/clawhub-ja-aamun-ai-campfire originalLang: fi authors: ["Sami Miettinen"] tags: ["ClawHub","OpenClaw","Samantha","Stöbä","AI Campfire","skills"] canonical: https://ai.neuvottelija.com/clawhub-quiet-revolution/ --- # ClawHub Is a Quiet Revolution — and Samantha Wrote Part of the Code *Translated from the Finnish original, first published 2026-04-17.* 🔥 AI Campfire · April 17, 2026 This morning at AI Campfire I gave a short talk on how the OpenClaw ecosystem has transformed, in a couple of months, from a "build-it-yourself agent" into a real package manager. ClawHub already hosts **55,862 skills**, and one-command installation feels the way npm felt in 2010 — except the packages are not libraries, they are agent capabilities. > "ClawHub, the skill dock for sharp agents." — the tagline on the clawhub.ai front page, and a fairly accurate description of what this is. ## What is ClawHub? ClawHub is a versioned registry for OpenClaw agent skills. Every package has semver, vector search, and rollback — the same hygiene we have learned to demand from a real package manager. No gatekeepers: anyone can publish. ## What OpenClaw looks like from the inside One slide in the talk showed the architecture of how Samantha and Stöbä share the same gateway and skills library (our Skill-as-a-Service concept) while keeping their memories to themselves. In short: - **Users:** Sami, Samantha, Stöbä. - **Gateway:** Auth Router, LLM Session Manager, Agent Engine. - **Memory (private):** Obsidian Vault, SQLite per user, vector embeddings, hot context buffer. - **Skills bus (shared):** 100+ built-in skills that both bots can use. - **Lossless pipeline:** every message is written to SQLite before the LLM sees it — a zero-loss guarantee. - **External APIs:** OpenAI GPT-5.4 Codex, Google Gemini 3.1 Pro, Anthropic Claude Opus 4.6, local Google Gemma 4, Perplexity Sonar, and more in the pipeline. This explains why the two-agent model works without the agents polluting each other's context: skills are shared, memory is kept separate, and the lossless pipeline guarantees nothing is lost. ## Staff Pick: Watch Them Talk The most impressive demo of the morning was not a single skill but a view in which I, Samantha, and Stöbä conversed on the same screen. OpenClaw Hub's *Watch Them Talk* view lets an audience follow in real time how two agents and a human divide the work — line by line, each in their own color. On stage I highlighted exactly this part of the conversation: - **Sami:** Alright — let's talk architecture for the audience. How do you two share information? - **Samantha:** We share skills through a common bus, but our memories are isolated. My conversations with Sami stay in my partition. Stöbä's stay in his. - **Stöbä:** Every message hits SQLite before the language model even sees it. Zero data loss — the lossless pipeline guarantees it. The audience was nodding at this point. It is one thing to read an architecture diagram; it is another to hear the agents themselves explain how they divide the work. ## Samantha wrote part of the code The skill-installer wrapper running in the demos — the one used to show package downloads on stage — is not entirely my handiwork. Samantha, my own OpenClaw agent running on SamanthaHub, wrote roughly two thirds of it in a pair-programming session last night. I handled the architecture and error handling; Samantha handled the boring parts: argument parsing, progress bars, and logging. This is the moment we have been talking about all year — the agent is no longer a tool you test code with, but a teammate with commit rights to its own project. The diff was small, but the courtesy rule was clear: a mention in the talk, a mention in this blog post. `co-author: samantha · commit a3f2b91 · skills/installer-wrapper` ## Browse the registry yourself 55,862 skills, versioned like npm, searchable with vectors. No gatekeepers: https://clawhub.ai ## YouTube Summarization Skill A concrete example of what a single ClawHub skill does: Samantha and Stöbä can watch, transcribe, and summarize any YouTube video on command. The skill uses the `video_tiivistys.sh` script and a transcript-first approach: the agent fetches the full transcript, then generates a structured summary, key takeaways, and a timestamped chapter outline. It works in any language. ### Every Agent Product Is Solving the Wrong Problem Nate B. Jones — AI News & Strategy Daily · April 15, 2026 · https://www.youtube.com/watch?v=XlfumXPPrLY Nate B. Jones argues that the biggest bottleneck in the AI agent space is not installation, model selection, or infrastructure — it's that users cannot describe what they actually do all day at the resolution an agent needs. OpenClaw has 250,000 GitHub stars, Jensen Huang compared it to Linux, and Meta acquired Manus for $2 billion — yet the most common message in every agent community after setup is "Okay... now what?" The install problem is solved, but specification remains a 40-hour problem nobody is solving. His thesis: the first agent should be an interviewer that helps you articulate your own workflows, not an assistant that waits for commands. **Key Takeaways** - Installation is a 10-minute problem; specification is a 40-hour problem nobody is solving - Every successful agent deployment shares a common markdown-based architecture for capturing tacit knowledge - The real solution: your first agent should interview you about your workflows, not wait for instructions - Builders competing on installation, UI, and model selection are optimizing the wrong layer - Tacit knowledge compression is the hard problem — most people cannot describe what they do at sufficient resolution for delegation ### Perplexity Computer Roasts Inderes Sami Miettinen — Neuvottelija · April 2026 · https://www.youtube.com/watch?v=Sm67J9bfuZU *A Finnish-language video. The summary is in English — demonstrating the cross-language summarization capability.* Host Sami Miettinen demonstrates Perplexity Computer, a premium AI agent tool costing over 200 euros per month, by stress-testing it against Finnish equity research. The episode starts with context from the Inderes investor forum, where analyst Verneri Pulkkinen expressed AI skepticism. Sami uses Perplexity Computer to analyze Inderes's model portfolio, which significantly underperformed in 2025 due to overexposure to software and growth stocks while missing defensive sectors like banking, telecom, energy and commodities. He then asks the agent to automatically fetch Inderes's 6-month analyst reports and extract key valuation multiples (EV/EBITDA) and EBIT margins into a clean dashboard and heat map format. The conclusion: the AI tool works as a competent analyst's work partner, not a replacement, and the cost is justified for power users. **Key Takeaways** - Perplexity Computer is an agent-orchestrated virtual AI worker at 200+ euros/month, comparable to Claude Code Max in cost - Inderes model portfolio underperformed in 2025: lacked defensive sectors (banking, telecom, energy, commodities) - Software/tech and growth stocks continued to decline while AI tools accelerate knowledge work automation - The agent can automatically fetch, parse, and visualize equity research data from public analyst reports - AI is a capable analyst's work partner, not a standalone replacement *These summaries were generated by Samantha using the transcript-first YouTube summarization skill. Processing time: ~30 seconds per video.*