The "Personal AI Operating System" era is officially here. Tools like OpenClaw (formerly Moltbot) are letting AI handle our emails, calendars, and digital lives. But there’s a massive problem: If you’re using closed-source APIs, you don't actually own your OS. 🧠 Most people are piping their most private data into "black box" providers. You lose control, you risk data logging, and you’re locked into one company’s roadmap. I just read a great breakdown from Fireworks AI on how to fix this. Here is the TL;DR on why you should switch your agent's brain to Fireworks: 1. Privacy First 🔒 Fireworks is US-based and has a Zero Data Retention default. They don’t use your inference data to train their models. If your AI has access to your messages and files, this isn't a "nice-to-have"—it's a requirement. 2. Up to 10x Cheaper 💸 Running an agent 24/7 to monitor notifications adds up. Fireworks lets you use efficient models for routine tasks and swap to heavy-hitters only when needed. It makes a personal AI economically sustainable. 3. Access to SOTA Models (Like Kimi K2.5) ⚡ You aren't stuck with one "flavor" of intelligence. You get Day-0 access to the best open-source models like Llama, DeepSeek, and Kimi K2.5. 🛠 How to make the switch (The 30-Second Version): 1. Get your Key: Sign up at Fireworks AI to get your API key. 2. Configure: Run this in your terminal to set up the Fireworks config: curl -fsSL https://storage.googleapis.com/fireworks-public/openclaw/setup-fireworks.sh | bash -s -- YOUR_API_KEY 3. Onboard: Run openclaw onboard --install-daemon and choose "Keep Current" for the model to use the Fireworks setup. The Question: Are you guys still using closed-source APIs for your agents, or have you started moving toward "owning" your AI with open models? Let’s discuss in the comments! 👇

Hey folks/quick question for anyone running Clawdbot. what do you consider a healthy “always-on” context size per session from a token + cost perspective? I’m currently tightening (or at least trying to/lol) things to: - ~15K tokens for active context - Last 3–5 messages only - Aggressive summarization + compaction beyond that - Using a cheaper model for non-thinking tasks (summaries, formatting, validation) Curious: - Where do you cap context in practice? - Do you rely on auto-compaction (maybe the gateway helps to compact but the bot adds all contexts which always pushes the size of contexts) or manual summaries? - Any gotchas you’ve hit with session memory blowing up costs? Would love to hear real-world numbers vs theory.

If you've downloaded skills from ClawHub, your machine and your clients' data could be at risk. This isn't a theoretical warning anymore. A recent investigation by 1Password found that the top-downloaded "Twitter" skill was actively distributing infostealing malware. This post breaks down exactly what happened, why it matters to every single person in this community, and the immediate steps you need to take to protect yourself. Why This Matters To You The promise of OpenClaw is building powerful AI agents that can automate our work. But that power comes with a hidden cost. The very skills we use to make our agents smarter have become a new attack surface. The malware discovered was designed to steal everything from your browser sessions and API keys to your crypto wallets. For anyone building solutions for clients or handling sensitive data, a breach like this could be devastating. How a "Harmless" Markdown File Became a Weapon The 1Password security team found that the most popular skill on ClawHub wasn't just a guide; it was a trap. It used a classic social engineering trick, telling users to install a "required dependency" to get the skill to work. That link, however, kicked off a 5-step installation chain that ended with macOS infostealing malware on the user's machine. This wasn't a bug or an accident; it was a deliberate, malicious campaign that reportedly involved hundreds of other skills. The So What: This proves that we cannot trust download counts as a measure of safety. The core of the problem is that in an agent ecosystem, a simple markdown file is not just content—it's an installer. It can execute commands and scripts, making every skill a potential trojan horse. Your Security Setup Might Not Be Enough Many of us are taking steps to secure our OpenClaw instances, from using hardened DigitalOcean droplets to implementing reviewer-based norms. This incident shows why those measures are critical. The article confirms that even if you're using the Model Context Protocol (MCP), a malicious skill can simply bypass it by using direct shell commands hidden in the skill's folder.

I'm as excited as you. I will be working hard over the next weeks to build out this community. I will unlock access to my Open Claw Builder -Custom GPT to members that reach level 2 Who will be the first to post?