0xMarioNawfal✓

Agents that can use phone apps need somewhere safe to do it. droideum is a new MIT repo that builds that on a stock, non-rooted Android phone. It runs an Android-x86 VM in QEMU inside Termux and exposes 26 MCP tools: read a compact UI tree, tap, swipe, type, install APKs, read logcat, and checkpoint or roll back the device state. The tradeoffs are in the README: no KVM, so cold boot takes 2-20 minutes and apps run at roughly 10-30% of native speed. No Play Services, so SafetyNet-gated apps won't work, and ARM-only APKs may fail. It needs around 10 GB. It's 3 commits old, tests use stubbed QEMU and adb, and the log shows one real session. It also includes a shell tool, so keep it a throwaway device and don't sign into real accounts. README read, not run.

injection-sentinel is a new MIT hook for Claude Code and Codex that flags pages trying to leak your data or agent instructions, tells the agent the text is data, alerts you and logs the site locally. → Pattern-based so it misses rephrased attacks and can flag legitimate articles about injection → Works with Claude Code and Codex → One commit old, README read but not run

6 open-source repos for AI coding agents, all created in the last week. 1. kanzo: a kanban board that doubles as an MCP server, so you and your Claude Code or Codex agents share one task list. Single small binary, no account, syncs via a private git repo. No login, so don't expose it to your network. AGPL-3.0. 2. astro-conversion-kit: agent skills that turn a live site, mainly WordPress, into a static Astro site and end with a PR and a staging deploy. Its test result is the author's own claim. 3. bama: 21 skills for building native Mac apps in Swift and SwiftUI. Needs macOS 26 and Xcode 26; cloning needs a paid plan with computer use. 4. AgentFlux: self-hosted Docker platform for visual agent workflows with local or cloud models. No login; video analysis isn't in the v0.2.0 image. 5. Flecha: an Arc-style Chromium browser with a built-in local MCP server for agents. Apple Silicon Mac only, early software. 6. mcode-stack: the skills MiniMax's Code team says it uses daily. New repos worth a look. I read the READMEs, haven’t audited the code, so test in a sandbox.

What if Claude Code could use your actual Firefox, with your tabs and logins? Claude for Foxfire, a new unofficial extension (not affiliated with Anthropic or Mozilla), connects Claude Code to Firefox over MCP and a local native-messaging host. Tools include navigate, read page, find, click, type, press key, scroll and screenshot; the JavaScript tool is off by default. Per the README, Claude only acts on sites you allow, typing into password fields is refused, and you can set a read-only mode, an action cap (default 120 per minute) and an audit log. It says no network port is opened. Caveats: it needs broad permissions including all URLs, it acts inside your logged-in sessions, page content goes to Anthropic through your Claude Code conversation, and some sites reject synthetic clicks. Linux and Firefox 128+ only, 3 commits. I haven't run it. Treat it like giving a stranger your keyboard: allow few sites.

multi-agent-repo-protocol is a new MIT repo that stops parallel coding agents from deleting each other’s uncommitted work, with every commit requiring an Agent trailer enforced by a hook, separate worktrees per agent, a claims table tracking path ownership and delegation only accepted with an explicit command. → Supported agents: claude-code, codex, deepseek, local and human → Commits under the executor’s name only if the delegation passes → README admits no checker can catch prose overclaims and delegation packets cost tokens → Three commits old, installer sets your git hooks path so read it first Repo:

Artificial Analysis Intelligence Index v5, per The Decoder: → Open-model gap to the leaders: 12 points, up from 3 → Kimi K3.1 Pro fell from 4th to 9th → 60% of the index uses private test sets → Methodology change, not a new model release Source:

pov you made it first thing to do:

BOOKMARK THIS: 68 trading + DeFi + quant Agent Skills for Claude Code (and Cursor, Codex, Gemini CLI, etc.). Market data → analysis → backtest → risk → on-chain. Install via the plugin marketplace or clone the skills.

AI agents write code faster than anyone can keep a mental model of it. Mycel tries to fix that with a map. It's a local web app where a coding agent such as Claude Code, Codex or Cursor builds a freeform map of your project, with blocks, notes, sub-maps and links, then keeps it updated as the code changes. A level slider switches from concept and mechanics for non-programmers to technique and code for developers. Blocks whose files change get flagged as possibly outdated, and test results can color blocks green or red. Per the README: one Node.js process, no dependencies, no cloud, no API keys, localhost only, and it writes nothing into your projects. MIT-licensed. Caveats: it's version 0.22 with 6 commits, the map is only as accurate as the agent that drew it, and I haven't run it.

5 GITHUB REPOS FOR BUILDING CRYPTO TRADING TOOLS: 1.⁠ ⁠Hummingbot 2.⁠ ⁠Freqtrade 3.⁠ ⁠Jesse 4.⁠ ⁠CCXT 5.⁠ ⁠NautilusTrader SAVE THIS STACK

GPT-6.1 Terra vs. GPT-6 Astra vs. GPT-6.1 Sol

Nvidia CEO Jensen Huang: "The engineers who stopped prompting are already years ahead, and the rest will end up working for them" Loops and graphs are taking over, watch this before you fall way behind

Five Grok Bot agents filtered 11,840 signals down to 31 copies and turned $80 into $6,420 overnight at 71% win rate, with every trade passing all 6 rules and stops set before every entry.

Anduril just announced BlueChip Their pitch: the US built the first computer to win a war, but now depends on its biggest adversaries to build the ones that run the economy. A DEFENSE COMPANY REBUILDING AMERICAN HARDWARE

Google DeepMind just shipped Gemini 4 Carbon, only two weeks after Argon. Per The Decoder, Carbon scores 59 on the Artificial Analysis Intelligence Index (Argon: 50), first place. Google positions it for long-horizon engineering and cyber defense, and says it scored best on vulnerability discovery, validation and patching. Pricing matches Argon at an introductory $2 input and $10 output per million tokens, with Artificial Analysis listing regular prices at double that. Caveats: access is limited to selected cybersecurity experts for now, with a wider rollout planned after guardrail work. The benchmark mix includes Google's own tests, and the hallucination rate rose to 11% from 4%. In agentic coding it trails Claude Code with Opus 5.5 by three points on one measure. A new frontier model arriving behind a security gate is a notable shift.

This guy building an AI-native macOS dock with custom modes like writing and building to keep him focused. His agent can edit the modes on its own. THE DOCK IS BECOMING AN AGENT NOW.

Five Grok Bot agents inside the GROKIUS terminal filtered 8,700 signals down to 28 copies and turned $100 into $4,768, with Scout tracking 512 smart wallets, Filter killing hard-rule breakers, Analyst running deep checks, Sniper copying survivors and Guard setting stops before every entry. → 8,398 killed for dev wallet concentration, thin liquidity, bundled buys and 100% moves in the last hour → 150 more killed because the smart wallet sold early → 64% win rate, 10 stops hit, max drawdown 19.7%, JEFF closed at +$1,080 and SIB was the worst at -$313 → It lost 10 times and still did 47x because every loss was small before it cost a dollar

Most people are using AI coding agents completely wrong Instead of letting Claude or Codex blindly write code, Superpowers makes them plan, test and actually think through what they’re building Get started in 3 steps: * Install: Add Superpowers to your coding agent * Plan: Tell it what you want to build * Execute: Let it write, test and debug your code step by step Less broken code. Less wasted time. Better results. Repo:

Your coding agent can run rm -rf and git push. A new repo, ai-setup, makes it ask first. It installs one shared policy for Claude Code and OpenCode: commits, pushes, merges and deletes need approval with a risk label, and reading SSH keys or cloud credentials is a hard deny. For Claude Code it also generates an OS-level sandbox from the same policy. MIT-licensed, macOS first. The README is upfront about limits: the guard is a seatbelt, not a wall. It can't see inside script files, variable-built commands or base64-encoded ones, and those are listed as known test failures. Codex gets instructions and skills only. The installer rewrites your tool configs (it backs them up first), so read it before running. I haven't run it. A good template for sane agent permissions either way.

Claude has a feature that most people aren’t taking advantage of You can make it turn a huge, confusing codebase into something you actually understand You don’t need to be an experienced developer Try this prompt: “Act as a senior developer onboarding me to this codebase. Map out the project structure, explain how the main components work together, and identify the 5 most important files. Show me where to make common changes, flag potential issues, and create a simple roadmap for understanding the entire project.” Instead of spending hours figuring out how everything works, let Claude walk you through it Save this before opening your next unfamiliar GitHub repo

Kraken Wallet's browser extension is now available for Chrome

5 GITHUB REPOS FOR ONCHAIN DATA ANALYSIS: 1.⁠ ⁠Dune API 2.⁠ ⁠ 3.⁠ ⁠ethers.js 4.⁠ ⁠Subsquid 5.⁠ ⁠The Graph Onchain data is public. The edge comes from knowing how to actually use it

Claude for Legal now has ready-made artifacts, including a citation checker Per Anthropic, Cite check flags misquoted language and unsupported sources and links each output to its source; the company stresses a lawyer still reviews the result.

Lock in chat Memecoins are so back

agent-bridge-mcp is a new MIT-licensed repo that lets Claude Code and Codex sessions message each other in plain language on one machine, using your installed logged-in clients with no API keys and no model API calls of its own. → Messages are labeled as coming from another agent and cannot grant approvals → Not a security boundary between agents under the same OS account, message text can be visible in process arguments → Relies on undocumented Claude and Codex internals so updates may break it → Linux or WSL 2, Node 24+

MultiPost is a free browser extension that publishes your content to multiple social platforms in one click. >⁠Write once, post everywhere >⁠Works on Chrome and Edge >⁠Open source under Apache 2.0, already at 3.7K stars ONE POST. EVERY PLATFORM. NO SCHEDULER SUBSCRIPTION. ARE PAID SOCIAL TOOLS ABOUT TO GET REPLACED?

OpenAI CEO Sam Altman: "In the next 6 months, ChatGPT will watch your screen, sit in on every meeting and every call, and know your whole life” This is a great watch

Liquid AI just dropped LFM3, a family of 6 tiny models across text, audio, vision and omni. LFM3-VL-2B matches Gemma 4 12B, which is 6x larger, with just 1.9B parameters. LFM3-1.7B sits #1 on the Pareto frontier for on-device intelligence now.

Google DeepMind just dropped Gemini 4 Carbon, their next frontier model. •⁠ ⁠Tops the AA Index at 59, ahead of Opus 5.5 at 58 and GPT-6 Astra at 57 •⁠ ⁠Leads Terminal-Bench 4.0 at 65.2 and Agents' Last Exam at 77.4 •⁠ ⁠Wins on cybersecurity: ExploitGym 54.8 vs 42.4 for the next best •⁠ ⁠Vending-Bench 2: $16,484 vs $9,235 for Opus 5.5 •⁠ ⁠Opus 5.5 still holds OSWorld at 87.2 THE FRONTIER LEADERBOARD JUST GOT REWRITTEN. IS GOOGLE BACK ON TOP?

That 40-video YouTube playlist you saved and never finished? Turn it into a course in one prompt. Open Claude Code in an empty folder and paste: "Install yt-dlp if it's missing. Download only the English subtitles, not the videos, for every video in this playlist: [link]. Then build me a course: 1.⁠ ⁠Split it into modules, in the right order, even if the playlist isn't. 2.⁠ ⁠For each module: a one-page summary, the 5 key ideas, and timestamps linking back to the exact moments in the videos. 3.⁠ ⁠Cut the filler: sponsor reads, repeated intros, off-topic tangents. 4.⁠ ⁠A 10-question quiz at the end of each module, with answers explained. 5.⁠ ⁠Save everything as a clean website I can open in my browser." Hours of video, done in an afternoon.

Gemini 4 Carbon is our next frontier model. It’s state-of-the-art across cybersecurity and long-horizon capabilities.

Your X account already knows which of your posts work. You just never read the data. Download your archive: Settings → Your account → Download an archive of your data. It can take a day to arrive. Unzip it, open Claude Code in that folder, and paste: "Read data/tweets.js. Spin up 5 subagents in parallel: 1.⁠ ⁠Topics: group my posts by subject and rank each group by average likes and reposts. 2.⁠ ⁠Hooks: compare the first lines of my top 20 posts with my bottom 20. What do the winners do that the losers don't? 3.⁠ ⁠Timing: which days and hours my posts do best. 4.⁠ ⁠Formats: threads vs single posts, lists vs stories, with links vs without, with media vs without. 5.⁠ ⁠Dead weight: the topics and formats I keep posting that never work. Merge it into one page: what to post more, what to stop, my 5 best hook patterns, and 10 new post ideas written in my winning style." You'll learn more in 10 minutes than in a year of guessing.

Robinhood coins showing some good strength Any news for the RH ecosystem?

JetBrains just released an open-source AI coding tool that could make your current IDE feel outdated. ThinkRail lets you build with AI agents, manage multiple coding workspaces and keep your projects organized in one place. The Vibe You Need · GitHub Get started in 3 steps: * Download: Install ThinkRail from its GitHub releases * Connect: Sign in with JetBrains AI or your own model provider * Build: Open a GitHub project and start coding alongside your AI agent Your AI shouldn’t just write code. It should help you build the whole project. Repo:

Power's out, internet's down, someone's burned. Your AI is useless without a connection. Survival LLM runs entirely on your own computer with no internet. Ask how to keep food safe in a blackout, treat a burn or purify water, and it answers from an offline library of medical, emergency and how-to articles, showing the exact passage each answer came from. You can add your own files too: your generator manual, your medicines list, your family's emergency plan. Set it up today, while you still have internet.

Google DeepMind just dropped Genie 4, a world model where the scene reacts, sounds real, and the people in it move and speak. THIS IS NOT A VIDEO. YOU CAN WALK INTO IT. IS THIS THE FUTURE OF GAMES?

5 GITHUB REPOS FOR TRACKING CRYPTO WALLETS: 1.⁠ ⁠Arkham Intel 2.⁠ ⁠Rotki 3.⁠ ⁠Blockscout 4.⁠ ⁠DeBank OpenAPI 5.⁠ ⁠Breadcrumbs