Open Source AI Agent OS in ~32MB File

Today’s top AI Highlights:

1. Cloudflare releases one API to crawl entire websites

2. Open-source AI Agent OS in one ~32MB binary

3. Open-source Hermes Agent that grows with you

4. Deterministic mock LLM server for testing

5. Build in parallel, design in context, & ship it with Replit Agent 4

& so much more!

Read time: 3 mins

AI Tutorial

How I Built an Autonomous AI Agent Team That Runs 24/7

Six AI agents run my entire life while I sleep.

Not a demo. Not a weekend project.

A real team that works 24/7, making sure I'm never behind. Research done. Content drafted. Code reviewed. Newsletter ready. By the time I open Telegram in the morning, they've already put in a full shift.

By the end of this, you will understand exactly how to build an autonomous AI agent team that runs while you sleep.

We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads) to support us!

Latest Developments

Cloudflare Releases One API to Crawl Entire Websites

The company that built one of the internet's most aggressive bot-blocking systems now has an official crawling API. Peak character development.

Cloudflare just shipped a /crawl endpoint (open beta) that lets you crawl an entire website with a single API call. Give it a starting URL and it discovers pages, renders them in a headless browser, and returns content as HTML, Markdown, or structured JSON.

The crawler respects robots.txt and Cloudflare's own AI Crawl Control by default, so it's designed to be the "right way" to crawl.

Built for RAG pipelines, training data, and content monitoring, with fully async with job IDs and polling. Available on both Free and Paid Workers plans.

Key Highlights:

1. Multiple output formats - Get crawled content as HTML, Markdown, or AI-extracted structured JSON via Workers AI, ready to feed directly into your LLM workflows.

2. Crawl scope controls - Set depth limits, page caps, and wildcard include/exclude patterns to target exactly the pages you care about.

3. Incremental crawling - modifiedSince and maxAge parameters skip unchanged pages, so repeat crawls don't waste time or credits.

4. Static mode - Set render: false to skip the headless browser entirely for a fast HTML fetch, ideal for documentation sites and blogs.

There are tons of optional parameters for more customization. URL sources, wildcard patterns, caching, and more. Check it out!

Open-source AI Agent OS in One ~32MB Binary

Move over, LangChain and CrewAI.

Someone looked at the Python agent framework ecosystem and said, "what if we just... didn't do that."

OpenFang is an open-source Agent OS written entirely in Rust - 137K lines of code that compiles down to a single 32MB binary.

No dependency bloat, no Docker pulls, no pip install chains.

OpenFang includes "Hands," autonomous agent packages that don't sit around waiting for your input. You activate a Hand, and it runs on a schedule: generating leads, monitoring competitors, managing your Twitter, cutting YouTube clips into shorts, or writing deep research reports.

7 Hands come pre-built, each with its own system prompt, skill files, and guardrails. The system supports 27 LLM providers, 40 messaging adapters, and has 16 discrete security layers, including a WASM sandbox and cryptographic audit trails.

Key Highlights:

1. Hands, not chatbots - Pre-built autonomous agents that run on schedules and deliver results (leads, research reports, social posts) to your dashboard or Telegram without any prompting.

2. Rust-powered performance - ~32MB install size, <200ms cold start, and 40MB idle memory, compare that to CrewAI's 3-second cold start and 200MB footprint.

3. 16-layer security model - WASM dual-metered sandbox, Merkle hash-chain audit, Ed25519 signed manifests, taint tracking, SSRF protection, and prompt injection scanning baked into the core.

4. One-command migration - Already on OpenClaw? openfang migrate --from openclaw imports your agents, memory, skills, and configs in one shot.

Open-source Hermes Agent that Grows With You 

OpenClaw set the standard for open-source personal agents that live on your machine and text you. Nous Research took that idea further.

Hermes Agent is their open-source autonomous agent that does the familiar stuff but comes with a twist: it's built by an AI research lab, and it doubles as infrastructure for training the next generation of tool-calling models.

Like OpenClaw, you install it on a server and chat with it via Telegram, WhatsApp, Slack, Discord, or CLI. It remembers your projects across sessions, spawns subagents, browses the web, and runs cron jobs in natural language. But Hermes is Python-native, which makes it immediately extensible for the ML crowd. And the same architecture powers Nous Research's RL training pipeline through Atropos, generating thousands of tool-calling trajectories in parallel for training models on agentic tasks.

Key Highlights:

1. Self-authoring skills - When Hermes solves a complex task, it writes a reusable skill document following the agentskills.io open standard.

2. Five sandboxed backends - Run commands locally, in Docker, via SSH, Singularity, or Modal cloud, all with container security hardening and persistent workspaces across sessions.

3. Built-in RL pipeline - Extends the Atropos framework with batch trajectory generation, 11 tool-call parsers, and trajectory compression, so you can train models on agentic behaviors, not just use them.

Quick Bites

Open-source mock LLM server for testing
CopilotKit just open-sourced LLMock, a zero-dependency mock LLM server that lets you test AI-powered apps without burning tokens or fighting non-deterministic outputs in CI. It spins up a real HTTP server on a real port, streams authentic SSE, and routes responses via JSON fixtures with substring, regex, or predicate matching. Especially handy for E2E setups where multiple processes (Next.js, Mastra, LangGraph) all need to hit the same mock endpoint.

Build in parallel, design in context, & ship it with Replit Agent 4
Replit just dropped Agent 4, and the big idea is parallel execution - multiple agents working on different parts of your project simultaneously, then merging results back together. It also collapses design and code into the same environment (infinite canvas, UI variant generation, direct-to-code edits) and lets you build web apps, mobile apps, slide decks, and data apps within a single project. Think less context-switching, more shipping.

Claude Code vs. Codex: A Dev's Honest Comparison
A developer who spent months on Claude Code, switched to Codex, and then switched back wrote a detailed comparison of the two, covering everything from Opus 4.6 vs. GPT-5.3-Codex task-completion benchmarks to token efficiency (Codex uses 3-4x fewer tokens) to a head-to-head RAG pipeline build. The verdict? Claude Code won on feel, ecosystem pull, and that $100/month mid-tier pricing. Worth the 12-minute read if you're deciding where to park your $200/month.

OpenRouter's two free stealth models are live
OpenRouter just dropped two free stealth models - Hunter Alpha (1T params, 1M context, built for agentic workflows) and Healer Alpha (multimodal with vision, audio, and real agentic execution). The internet's already deep in detective mode, leading theories point to DeepSeek V4 or another Chinese lab, with Hunter's reasoning traces leaking Chinese legal compliance reminders and Healer literally telling users it was "built by Xiaomi." Early tests show Hunter shining at tool use and long-horizon planning while Healer brings creative multimodal chops. Both available now!

Tools of the Trade

1. Agent Browser Protocol: A custom Chromium build that turns web browsing into discrete, atomic steps. Each API call injects input, waits for the page to settle, captures a screenshot, and freezes JS until the agent's next move. It ships with MCP + REST baked in, hits ~100ms overhead per action, and scores 90.53% on Online Mind2Web.

2. alphaXiv SKILL.md: Drop a file into Claude Code for structured paper lookups via API. No more parsing raw PDFs. First "killer skill" from a third party for the Claude Code ecosystem.

3. DeerFlow by ByteDance: Open-source "super agent harness" built on LangGraph and LangChain. It gives AI agents a full execution environment with sandboxed containers, persistent memory, a filesystem, and extensible skills so they can actually research, code, and produce deliverables.

4. Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
   (Now accepting GitHub sponsorships)

Hot Takes

1. As a founder, nothing is more validating than watching a competitor build a sloppier copy of your product and still get a 10 figure valuation.

   ~ Pietro Schirano
   
   

2. btw if you can build a category leader open source project in ai engineering right now the market acquihire rate is ~$10-$100m per ai engineer.

   you do not need to figure out a business model, you do not need GTM, you do not need funding.

   just build things clankers want.

   ~ swyx

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉