Generalizing problems is as important as generalizing solutions. Because once you generalize problems, your solution is automatically generalized. #programming

We're releasing AgentVFS, a virtual filesystem interface for AI agents agentvfs.org/ AgentVFS unifies every service into one interface and replaces the need of calling MCP. Stored data becomes one virtual filesystem across services, described by VFS.md, and actions become virtual CLIs, so the agent no longer calls MCP tools at all. AI agents work on live data across every service through the terminal-like interface they know best. They use the Unix tools they learned in training (ls, grep, jq, pipes, loops) and the terminal workflows they practiced in post-training on agentic coding. We're also releasing VFS-Bench 1.0 agentvfs.org/benchmark/vfs-b…, the first systematic study and benchmark to compare a virtual filesystem with MCP on the same tasks. Compared with MCP, AgentVFS uses up to 88% fewer tokens and has about a 5% higher success rate on average.

34 AI companies give away free API tokens every month. Stacked together, that's about 7.4 billion tokens. FreeLLMAPI puts all of them behind one OpenAI-compatible endpoint. 37.8K stars. → add your free keys from Google, Groq, Cerebras, Mistral, OpenRouter, Cloudflare, NVIDIA, Hugging Face and more → it picks the best available model for each request → when one provider hits its limit, it switches to the next one automatically → tracks usage per key so you stay under every free cap → your keys are stored encrypted → works with any OpenAI-compatible app, plus coding agents like Codex, Cursor, Aider and Claude Code Install (needs Docker): curl -fsSL freellmapi.co/install.sh | bash Or grab the Mac or Windows desktop app from the repo's Releases. Built for personal projects, not production: free tiers have no top frontier models and get slower late in the day. MIT.

Terminal EPUB/PDF/DJVU reader focused on speed, smooth navigation, and Vim-style workflows. github.com/bugzmanov/bookokr…

Don't waste all your time improving how your AI agents reason. Do this instead... Spend more time defining what they should reason about. Even the smartest model can make a bad decision if the options are poorly defined. This is why I like using Jev and ontologies together... Jev by @typesafeai turns unstructured inputs into structured decisions with calibrated confidence scores. Instead of asking an LLM to reason through every decision, you define the possible outcomes in code. Jev selects an outcome and returns its confidence. And the ontology defines what those outcomes mean. Imagine an AI agent handling customer support tickets. The ontology defines: • Which issues belong to billing versus technical support. • How customers, subscriptions, and invoices relate. • Which categories should never overlap. Jev classifies tickets against those categories. Based on its confidence, your agent can: 1. High confidence → Execute an action. 2. Medium confidence → Escalate to an LLM. 3. Low confidence → Request human review. So expensive reasoning is reserved for cases that need it. But what if your ontology is wrong? Jev can confidently choose the wrong category if the options are poorly defined. You still need evals to identify failures, improve definitions, and test changes before deploying. As those definitions evolve, version them like your prompts, datasets, and code. The takeaway: Building better AI agents isn't always about making the model smarter. Sometimes it's about defining its world more clearly. P.S. How are you defining the decision space for your AI agents?

and now my redesign is open source too. it's called harp github.com/dominikmartn/harp

MindOS is a clean, fluid operating system built on the Linux kernel. Its goal is to remove the barriers between you and your computer: it boots fast, offers a lightweight, bloat-free interface and feels like a system that truly belongs to you. mindos.studio/

You can now run Google's Gemma 4, a 26-billion-parameter AI model, on a basic 8 GB MacBook Air. Free, offline, no subscription. The model normally needs 14.3 GB of memory. TurboFieldfare keeps a small 1.35 GB core in memory and pulls only the parts each token needs from your SSD, so it runs in about 2 GB of RAM. → about 5 to 6 tokens per second on an 8 GB M2 MacBook Air, and 31 to 35 on an M5 Pro → native Mac app: click Download Model, then Load Model, then start typing → also comes with a command line and a local OpenAI-compatible server, so other apps can use it Needs an Apple Silicon Mac on macOS 26, Apple's Swift tools, and about 15 GB of free disk space. git clone github.com/drumih/turbo-fiel… cd turbo-fieldfare swift build -c release .build/release/TurboFieldfareMac Apache-2.0, 6.9K stars.

LordOS is a free GNU/Linux distribution based on Debian with its own desktop environment, display server and Wayland compositor written in C, its own interface engine, and native apps made to feel calm, clear and private. lordos.org/

If you want to integrate Skillspector into your agent or project, just give this skill to Claude Code github.com/NVIDIA/SkillSpect… It's worth scanning your skills, especially if you've been installing random ones you find on X

Agent tree component by @flohoeller

You can now run 290B-parameter open-source AI models on a gaming PC. FreeToken is a free engine that runs frontier-size open models like DeepSeek V4 Flash, Qwen and GLM-5.2 on hardware you already own. 14.6K stars. → splits the work across your GPU, CPU and system memory instead of needing a datacenter → keeps the most-used parts of the model in fast memory and swaps the rest in as needed → works with NVIDIA RTX 30, 40 and 50 series cards → OpenAI- and Anthropic-compatible API, so tools like Codex, Claude Code and OpenCode can run on your own machine → no per-token bill, and nothing leaves your computer Easiest way: download the desktop app for Windows or Linux at flashml.ai. It sets everything up and gives you a chat window. Or with Python: uv pip install "freetoken[accel]" Apache-2.0. github.com/FlashML-org/FreeT…

DeepSeek Bot (DSH + Memory) github.com/rlaope/deepseek-b…

Must-read papers of the week ▪️ Cortex ▪️ Librarian Agents ▪️ OpenGCUA ▪️ DuoDiscover ▪️ WAVE-0 model ▪️ ACT3D model ▪️ Execution Before Memory ▪️ Has GPT-6-Astra Achieved Embodied Intelligence? ▪️ Persistent Planar Memory ▪️ Relational Learning as World Modeling ▪️ More Methods, Better Agents? ▪️ Mosaic: Multi-Session Agentic Reinforcement Learning ▪️ HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents ▪️ Into the Unknown: Benchmarking AI agents on open scientific research problems Explore them to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest: turingpost.com/p/better-ai-a…

Microsoft now ships Linux commands that run natively on Windows. Coreutils for Windows brings ls, cp, cat, grep, find and the rest of the classic Linux toolkit to Windows, so the same commands and scripts work on Windows, Mac, Linux and WSL. 5.3K stars. → one install: winget install Microsoft.Coreutils → the same flags and pipelines you already use on Linux → built on the open-source uutils project, packaged and maintained by Microsoft → switch off any command that clashes with your shell Still in preview. PowerShell needs version 7.4 or later, and a few commands like kill aren't included. MIT. github.com/microsoft/coreuti…

Want a desktop AI agent that can run local models instead of defaulting every task to a cloud API? ZhiYuan Agent is an open-source, local-first desktop AI agent for builders who want a workspace for agent tasks and local GGUF models. It helps you move from a task description to an observable workflow by showing progress and approval requests while its tools work with files, commands, and the browser. Key features: • Built-in free model – start without a third-party API key; the README notes that network access and usage limits apply • Local GGUF models – choose a model and quantization in its marketplace, then tune context, GPU allocation, and threads through the llama.cpp runtime • Tool-using workspace – agents can read and write files, run commands, and operate a browser • Chat-app task intake – assign tasks from WeChat, WeCom, DingTalk, Feishu/Lark, QQ, or email • Approval controls – operations can request approval, with progress and results kept visible in the workspace It’s open-source (GNU Affero General Public License v3.0 license). Link in the reply 👇

At @aiDotEngineer, I presented our skill called /skill-doctor. It's a skill to improve your agent skills. It aggregates your past agent sessions, grades their quality, and propose improvements to fix common failures.

Building an agent skill is one thing. Knowing when it actually triggers is the harder part. OpenCode Skill Creator is a workflow skill and TypeScript plugin for creating, evaluating, and improving OpenCode skills. It helps you move from a draft to a tested skill by generating trigger evaluations, measuring description accuracy, and comparing task outputs before installation. Key features: • Skill drafting – creates a SKILL.md workflow with frontmatter and supporting files • Trigger evaluation – builds should-trigger and should-not-trigger queries to measure skill selection • Description optimization – improves descriptions using eval failures and a train/test split • Run comparison – aggregates graded outputs, timing, and token usage across runs • HTML review – lets you inspect paired outputs in a browser before installing a finished skill It’s open-source (Apache License 2.0). Link in the reply 👇

Speed up your code like crazy! Maxwell is a new kind of AI. It's a code execution model that speeds up your code on CPUs and GPUs. "Code execution models" are a new category of AI. Maxwell is the first one, but this category will explode in popularity, and you'll see why. Say you have a script that moves and processes some data. You integrate your code with Maxwell, run it, and let it observe how it runs on your hardware and analyze the data your script moves around. Maxwell will identify bottlenecks and look for faster algorithms and machine instructions for your workload. As long as your script runs, Maxwell rewrites the binary to make it faster. In one example, Maxwell made a pandas pipeline 38.4x faster. It used more CPU cores, combined processing steps, reduced unnecessary memory reads and writes, and selected the right algorithms for the data the script was processing. For example, imagine you want to check whether a number belongs to a collection. You could use hashing, but if the numbers occupy a small, dense range, a compact bitmap can make that check much faster. In short, that's one of the algorithmic data-dependent tricks Maxwell can leverage to speed things up. Here is the most impressive part of all of this: Maxwell produces a machine-checked mathematical proof that each replacement preserves the original behavior. Basically, any code they modify will always produce the same output as the old one. This is really cool for anyone working on AI or data engineering workflows. You will squeeze much more work from the hardware without rewriting your application. You can try out your scripts using Maxwell by going to this link: fandf.co/4iugTRP Thanks to the @_doubleAI_ team for partnering with me on this post.

Your AI coding agents shouldn’t relearn the team’s context from scratch. Cortex Hub is a self-hosted AI agent memory and code-intelligence platform for teams using AI coding agents across projects and IDEs. It helps you carry decisions, fixes, and code context between agents by putting persistent memory, shared knowledge, and AST-aware search behind one MCP endpoint. Key features: • Persistent agent memory – stores and semantically recalls context across sessions, projects, and branches • Cross-project code search – searches indexed repositories by symbol or flow without requiring a separate repo lookup • Shared knowledge base – records team findings with temporal validity and supersession links • Code-impact tools – traces callers, callees, and downstream impact before an edit • Quality enforcement – includes session scoring plus pre-commit gates for build, regression, standards, and traceability checks It’s open-source (MIT license). Link in the reply 👇

There's now a free, open-source alternative to Microsoft Office with an AI agent built into every document. GenOffice opens and saves real .docx, .xlsx and .pptx files, plus PDF, Markdown and HTML: → only what you edit gets rewritten, so files still open cleanly in Word, Excel and PowerPoint → AI edits show up as tracked changes you can roll back in one click → spreadsheets get real formulas, not pasted numbers → convert PDFs to Word, Excel or PowerPoint on your own computer → search your files by what's inside them, and ask questions across documents with a source for every answer → sign in with Genspark, or use your own key for OpenAI, Claude, Gemini, DeepSeek, Grok and more, including local models Your files stay on your machine. Only the AI requests go to the provider you choose. Free for Mac, Windows and Linux. Apache-2.0, 10K stars. github.com/genspark-ai/genof…