Agents learn to see inside binaries, manage memory, and extend themselves this week
Oct 12, 2026 – Oct 18, 2026
This week's trending repos center on AI coding agents gaining new capabilities: reverse engineering, persistent memory, skills marketplaces, and orchestration across multiple harnesses. A reverse-engineering MCP server drew significant overseas attention, while tools for taming context windows, routing between agents, and extending Claude Code and Cursor with plugins also gained traction.
Agents get the ability to reverse engineer and inspect binaries
morluto/rea plugs AI agents like Claude Code and Codex into reverse-engineering tools via MCP, letting them explain binary behavior directly; it drew extensive overseas YouTube and Hacker News coverage and reached the top rank this week. The underlying NSA-built disassembly framework NationalSecurityAgency/ghidra it pairs with also trended, with coverage noting it keeps receiving new releases. Separately, vxcontrol/pentagi applies a similar multi-agent approach to penetration testing, running autonomous agents through a self-hosted system aimed at security researchers.
Featured repositories (3)
A single MCP server that lets AI agents reverse engineer anything — REA plugs into agents like Claude Code and Codex via MCP, pairing them with tools like Ghidra and Hopper to decompile binaries and trace app behavior. Install with one npx command and let an agent explain how a feature works down to the binary level.
Reverse engineer anything with agents, from app behavior down to native binaries.
NSA's reverse engineering suite keeps evolving with new releases — Ghidra is the NSA-built framework for disassembling, decompiling, and scripting analysis of compiled binaries across Windows, macOS, and Linux. It supports extensions in Java or Python and both interactive and automated workflows, making it useful for malware analysts and vulnerability researchers. Check current security advisories before deploying a given release.
Ghidra is a software reverse engineering (SRE) framework
Self-hosted multi-agent AI system for automated penetration testing — PentAGI runs multiple autonomous AI agents, built with Go and a React frontend, to carry out penetration testing tasks end-to-end. It supports models like GPT and Anthropic, exposes a GraphQL API, and targets security researchers and ethical hackers wanting self-hosted, automated offensive security workflows.
Fully autonomous AI Agents system capable of performing complex penetration testing tasks
Managing context and memory across long agent sessions
Several repos address agents losing track of state over long sessions. mksglu/context-mode sandboxes tool output and persists session memory across coding-agent platforms. akitaonrails/ai-memory goes further, letting agents hand off a task to a different harness mid-session while retaining progress and open questions. JuliusBrussee/caveman takes a token-reduction approach, stripping verbose agent output to cut costs. xerj-org/xerj indexes code, docs and logs locally so agents can search and reuse existing solutions instead of re-reading files.
Featured repositories (4)
Taming bloated context windows in AI coding agents — Context Mode sandboxes tool output to cut context usage by 98%, persists session memory across runs, and enforces routing rules across 17 coding-agent platforms via MCP and hooks. Built for teams running Claude Code, Codex, Copilot, Cursor and similar tools who need agents to stay efficient over long sessions.
Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks.
Shared memory that lets coding agents swap places mid-task — ai-memory is a Rust tool giving AI coding agents a shared, cross-vendor memory. It lets you quit one harness (Claude Code, Codex, Cursor, Gemini CLI, and 20+ others) and resume in another with full context on progress, failures, and open questions, plus a portable profile of your work habits across projects.
Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
Caveman shrinks coding-agent tokens by talking like a caveman — Caveman is a Go-based skill and proxy for coding agents like Claude Code that strips verbose output to cut token use, citing 33% fewer input tokens and up to 54% cost savings across 30+ agents. It's backed by research mentions from Adobe and JetBrains, aimed at devs running expensive LLM coding sessions.
🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
A single Rust binary that gives AI coding agents real search — XERJ indexes code, docs, logs and PDFs locally for BM25, vector and hybrid search behind an Elasticsearch-compatible API. Its reference-coding feature lets agents like Claude Code or Cursor find and reuse solutions from similar open-source repos instead of re-reading files, cutting output tokens by up to 2.7x in testing.
AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.
Skills and plugin marketplaces extend agents into new domains
Anthropic published its own curated plugin directory for Claude Code, anthropics/claude-plugins-official, covered by Japanese tech media, giving a vetted way to add commands, agents and MCP servers. Cursor shipped a comparable plugin specification and registry, cursor/plugins. Domain-specific skill sets also trended: K-Dense-AI/scientific-agent-skills bundles validated skills and scientific databases for research tasks, android/skills grounds agents in Android development patterns, and cathrynlavery/diagram-design generates brand-matched diagrams without requiring Mermaid syntax.
Featured repositories (5)
Anthropic's own marketplace for Claude Code plugins — This is Anthropic's official, curated directory of Claude Code plugins, split into internal (Anthropic-built) and external (partner/community) entries. Plugins install directly via Claude Code's `/plugin install` command, giving users and third-party developers a vetted, standardized way to extend Claude Code with commands, agents, skills, and MCP servers.
Official, Anthropic-managed directory of high quality Claude Code Plugins.
Official spec and plugin registry for extending Cursor agents — Defines the manifest format for Cursor plugins and ships a growing collection of official ones, from CI workflows and PR review canvases to self-referential AI loops and repo compatibility scanners. Useful for developers who want to extend Cursor's agent behavior or build their own plugins.
Cursor plugin specification and official plugins
A rebranded, open-standard skills library turning agents into science assistants — Formerly Claude Scientific Skills, this repo now works with any Agent Skills-compatible tool (Cursor, Codex, Pi, Antigravity), not just Claude. It bundles 177 validated skills and 100+ scientific databases for biology, chemistry, medicine, and drug discovery, and powers the new open-source desktop co-scientist K-Dense BYOK.
Turn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 250,000+ scientists worldwide. 177 ready-to-use validated skills plus 100+ scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.
Official AI agent skills for Android development gaps — A collection of markdown-based SKILL.md files following the open agent-skills standard, designed to ground LLMs in Android development patterns where they currently underperform, like R8 analysis or Wear Compose. Installable via the Android CLI for agents like Gemini, Antigravity, or Claude Code plugins.
An agent skill that draws brand-matched diagrams, no Mermaid required — This is an agent skill for coding assistants like Claude Code, Codex, and Cursor that generates diagrams as self-contained HTML+SVG files across 44 types. It can pull a site's palette and fonts to match your brand, and imports from drawio, Mermaid, or Excalidraw, aiming for editorial-quality output rather than default auto-layout look.
Editorial diagram design for Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, Qoder, and Pi. 45 diagram types. Self-contained HTML + SVG. No shadows. No Mermaid slop.
Orchestrating and routing across multiple coding agents at once
Rather than building a new agent, several projects coordinate existing ones. stablyai/orca runs several coding agents in parallel git worktrees from one desktop app with mobile notifications. Alishahryar1/free-claude-code routes across many free-tier providers so coding agents keep working during outages. ruvnet/ruflo extends multi-agent swarm tooling with adaptive memory and federation, while affaan-m/ECC layers skills, instincts and persistent memory on top of existing agent harnesses rather than replacing them. HKUDS/CLI-Anything takes a different angle, auto-generating command-line wrappers so agents can drive ordinary software.
Featured repositories (5)
An orchestrator that runs Codex, Claude Code, OpenCode and Pi side-by-side — Orca lets you run multiple coding agents (Codex, ClaudeCode, OpenCode, Pi) in parallel, each in its own git worktree, tracked from one desktop app. A mobile companion app notifies you and lets you steer agents remotely. Built for builders juggling several AI coding agents with their own subscriptions.
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
A free-model router that lets 11 coding agents share one catalog — FCC connects Claude Code, Codex, Aider, Cline and other agents to 61 ToS-friendly providers offering 1.3B+ free tokens monthly. It auto-switches models during outages, trims token usage via RTK/Headroom, and adds browser-based sessions, useful for anyone juggling multiple AI coding tools without risking accounts.
Use Claude Code, Codex, Antigravity, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)
A meta-harness evolving Claude Flow's multi-agent swarm tooling — Ruflo is an agent meta-harness for Claude Code and Codex, tracing its lineage to the original Claude Flow project. It adds adaptive memory, self-learning behavior, federation, and vector RAG support, installable as Claude Code plugins for developers building multi-agent swarms and conversational AI systems.
🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
A performance layer that gives coding agents skills and memory — ECC adds skills, instincts, persistent memory, security checks and research-first workflows on top of Claude Code, Codex, Opencode, Cursor and similar agent harnesses. It targets developers who want more reliable, context-aware AI coding agents rather than a brand-new harness itself.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Turning any software into a command-line interface AI agents can drive — CLI-Anything auto-generates CLI wrappers so AI agents can operate existing software without GUIs. Its companion CLI-Hub lets you install and share community-built CLIs via pip, aiming to make everyday tools usable by autonomous agents rather than just humans.
"CLI-Anything: Making ALL Software Agent-Native" -- CLI-Hub: https://clianything.cc/
Beyond agents: creative tools and platform ports also drew attention
Outside the agent ecosystem, storytold/artcraft lets artists stage scenes before generating AI images or video, aiming for repeatable composition rather than raw prompting. bilawalsidhu/gods-eye-view turns public flight, ship and satellite feeds into a live, voice-controlled globe view. boykopovar/AnyPS5 relinks PS5 executables into native Linux and Windows binaries without emulation, and debpalash/VoiceStudio bundles voice cloning, dubbing and transcription into a local, open-source studio covering a wide range of languages.
Featured repositories (4)
An IDE that lets artists stage scenes in 3D before generating AI images — ArtCraft is a desktop IDE for AI image and video creation, combining 2D composition with 3D scene staging so artists can control composition, character identity, and camera placement before generating. Built in Rust, it targets artists, designers, and filmmakers who want repeatable, precise results instead of raw prompting.
ArtCraft is an intentional crafting engine for artists, designers, and filmmakers
Open source globe turns public flight, ship and satellite feeds into a live spy-sat view — This Cesium-based 3D globe streams real aircraft, ships, satellites, earthquakes, traffic and public cameras, plus hands-free voice control via a realtime AI agent. Spun out of the viral YouTube/WorldView series, it runs without API keys via Pinokio or locally, aimed at OSINT and geospatial hobbyists.
A spy satellite simulator in your browser, except the data is real. Live open source spatial intelligence on a photorealistic 3D globe.
Porting PS5 executables to Linux and Windows without emulation — AnyPS5 relinks PS5 executables into native Linux/Windows binaries, reimplementing system prx libraries instead of emulating them, and recompiles shaders to SPIR-V. It has a browser-based converter, controller/keyboard input mapping, and a community-tracked per-game compatibility map, aimed at enthusiasts exploring native PS5 game ports.
Tool for automatic PS5 executables porting to Linux and Windows
A fully local, open-source voice studio covering 646 languages — VoiceStudio bundles voice cloning, voice design, dubbing, dictation, transcription and audiobook creation into one local-first tool, positioned as an open-source ElevenLabs alternative. It runs on CUDA/MLX, exposes a local API and MCP for agents, and supports watch folders and reusable recipes for automated workflows.
VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
The week in numbers
- Repositories on daily trending (all languages): 38
- Total stars gained on listed days: +275,492
- Categories: AI agents 22, LLMs & ML 5, Apps & games 4, Developer tools 3, Security 2, Data & databases 1, Other 1
Buzzing topics this week
Hacker News stories counted by topic for this week (Monday to Sunday). Shown from our own tally, separately from the article above.
AI
SteadySecurity
Rising- Fingerprints are the new cookies ↗ · Hacker News
- Yandex bot traffic down 99% after data center attack ↗ · Hacker News
- I Found a $90k Remote Code Execution Vulnerability in Google ↗ · Hacker News
LLMs
Steady- Muse Spark Open ↗ · Hacker News
- The gap in shared understanding of LLM capability is widening ↗ · Hacker News
- GLM-Legal ↗ · Hacker News
PCs & hardware
Steady- Whatever became of holography? (2011) ↗ · Hacker News
- Turn a WiFi 7 Router into a 320Msps Software Defined Radio Receiver ↗ · Hacker News
- Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X ↗ · Hacker News
Open source
Steady- Doory – Open-Source Canva ↗ · Hacker News
- OpenFatigue™ – Open is the new closed ↗ · Hacker News
- Iosevka is an open-source, sans-serif and slab-serif monospace font ↗ · Hacker News