10 Fastest-Growing GitHub Repos
Verified Deep-Dive & Security Audit

The one-screen verdict

† Medium when deployed as documented (Docker/VPS + allowlists); High if installed carelessly on your main machine. Stars fetched live from the GitHub API on 2026-07-19.

What it is

An Agent Skill (Claude Code, Codex, Cursor, Gemini CLI…) that adds a /last30days <topic> command. It fan-out searches ~20 sources — Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub, arXiv — for the last 30 days, ranks results by real engagement (upvotes, likes, prediction-market money), and synthesizes one cited brief. Reddit/HN/Polymarket/GitHub work with zero API keys.

What's inside

• skills/last30days/SKILL.md — a 2,148-line runtime spec the agent follows (setup wizard, flows, flags)
• scripts/last30days.py + ~100 modules in scripts/lib/ — per-source clients, pipeline, fusion, rendering
• store.py / watchlist.py / briefing.py — recurring trend monitoring with Slack/webhook delivery
• mcp/ — Go MCP server builds for Claude Desktop; multi-host plugin manifests for Claude/Codex/Grok

Install & use

/plugin marketplace add mvanhorn/last30days-skill    # Claude Code
/plugin install last30days
npx skills add mvanhorn/last30days-skill -g           # Codex / Cursor / Gemini
/last30days claude code skills                        # then just ask

How you can use it

• Pre-meeting dossiers on a person/company: recent X posts, GitHub activity, podcast mentions
• Competitive comparisons with live stars + community sentiment (X vs Y, --competitors)
• Content research: what Reddit/X actually says about a niche this month, before you write
• Per-client scheduled trend monitoring via --store + watchlists

Security

• Secrets handled well: keys in ~/.config/last30days/.env written 0600, or macOS Keychain; no eval/shell=True; XSS-hardened HTML output; OpenSSF Scorecard CI.
• Caution: keyless web fetch proxies URLs through r.jina.ai (third party sees your queries); optional publishing uploads briefs to api.ht-ml.app; the setup wizard executes a pinned npm package via npx; it can (with consent) decrypt X.com cookies from your local Chrome profile.
• Prompt injection gap: its whole job is piping untrusted social-media text into your agent's context — and SKILL.md contains no injection warning or mitigations.

Verdict: unusually well-engineered for a viral skill; install it, but use throwaway/scoped API keys and decline browser-cookie access unless you need X results.

What it is

A local "context compression layer" that sits between your AI agent (Claude Code, Codex, Cursor…) and the LLM API, shrinking tool outputs, logs, JSON and history before they're sent — claiming 60–95% fewer tokens on JSON-heavy content with the same answers. Compression is reversible: originals are cached locally and the model can query them back.

What's inside

• headroom/ — Python library: SmartCrusher (JSON), AST CodeCompressor (6 languages), Kompress-v2 ML text compressor
• headroom/proxy/ — FastAPI localhost reverse proxy with ~100 policy modules (auth, redaction, caching)
• headroom/ccr/ — reversible-compression store + MCP server; crates/ — Rust core rewrite
• TypeScript SDK, Claude Code hooks, opencode/openclaw plugins, benchmark suite

Install & use

uv tool install --python 3.13 "headroom-ai[all]"
headroom wrap claude        # launches Claude Code through the compression proxy
headroom doctor && headroom dashboard
headroom unwrap claude      # undo

How you can use it

• Cut Claude Code / Codex API bills (~15–20% typical, more on JSON-heavy tool output) with one command
• Compress RAG chunks / scraper output inline via compress(messages) in your pipelines
• Fit longer working sessions into the context window before compaction kicks in

Security

• By design it MITMs all your LLM traffic (localhost only). Auth headers are redacted from debug captures; telemetry is local-only and the old phone-home beacon was removed; an HEADROOM_OFFLINE air-gap switch exists.
• Caution: cached originals live unencrypted in ~/.headroom/ccr_store.db — every tool output your agent ever produced, in plaintext SQLite. Treat that folder as sensitive.
• Caution: loads HuggingFace tokenizers with trust_remote_code=True, and [all] pulls 269 locked packages (torch, transformers…) — a big supply-chain surface. Maintainer is attentive (proactively excluded a compromised dependency release).

Verdict: real savings for heavy Claude Code/Codex users; adopt it, skip [all] extras you don't need, and protect ~/.headroom.

What it is

Nous Research's open-source "self-improving" personal AI agent — the biggest repo on this list. A Python agent loop with a terminal UI plus a gateway that lets you talk to it from Telegram, Discord, Slack, WhatsApp, Signal, email and ~25 other channels. It writes its own skills from experience, keeps long-term memory about you, runs cron automations, and executes shell commands on local machine, Docker, SSH or cloud sandboxes. Model-agnostic.

What's inside

• agent/ — core loop: context engine, memory manager, skill creation/curation, secret redaction
• gateway/ + plugins/platforms/ — ~29 messaging connectors, DM pairing, webhooks, API server
• tools/ — 40+ tools; six terminal backends (local, Docker, SSH, Singularity, Modal, Daytona)
• skills/ — bundled skill library + skills_guard.py scanner; cron scheduler; desktop app + TUI

Install & use

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash   # install.ps1 on Windows
hermes setup && hermes                         # wizard, then terminal chat
hermes gateway setup && hermes gateway start   # then message your bot on Telegram

How you can use it

• Always-on assistant on a cheap VPS you message from Telegram — research, drafting, file wrangling
• Scheduled automations: daily client reports, backup checks, weekly repo audits delivered to Slack/Discord
• A trainable "ops person" that builds reusable skills for your recurring workflows

Security

• Better hygiene than most in its class: credentials 0600 in ~/.hermes/auth.json, env scrubbed from subprocesses, unknown chat users must pair with owner-approved codes, HTTP surfaces default to loopback, no first-party telemetry (it even disables a dependency's PostHog).
• Caution: the default terminal backend is local — LLM-generated shell commands run directly on your machine, and internet-facing chat channels feed it input. This is the same exposure class that produced real-world OpenClaw incidents; Hermes documents it honestly and ships containment options, but the defaults are permissive.
• Caution: skills execute arbitrary Python at import time, and scanning of agent-created skills (skills.guard_agent_created) is off by default. Optional cloud memory (Honcho/mem0) ships your user model to third parties.

Verdict: genuinely capable and honest about its trust model — run it in Docker or on a disposable VPS with allowlists, never bare on your main machine. † High risk if installed carelessly.

What it is

A pack of 13 portable Agent Skills — pure markdown instruction files that steer Claude Code / Codex / Cursor away from generic "AI slop" frontends. They encode opinionated design rules: typography, layout variance, motion direction, banned clichés (centered hero, purple gradients, Inter-everywhere). No runtime code at all.

What's inside

• design-taste-frontend (v2, 1,207 lines) — flagship: brief inference, VARIANCE/MOTION/DENSITY dials, design-system mapping, canonical GSAP ScrollTrigger skeletons, pre-flight checks
• gpt-taste — stricter GPT/Codex variant; redesign-skill — audit-first refresh of existing UIs
• soft-skill / minimalist-skill / brutalist-skill — fixed aesthetic directions
• output-skill — anti-truncation (bans // TODO placeholders); imagegen-web/mobile, brandkit — image-prompt skills for mockup boards

Install & use

npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"
# then prompt normally — the skill auto-triggers, states a "Design Read", ships code

How you can use it

• Drop into Vite/React client projects so landing pages stop looking AI-generated — its GSAP pin/scrub skeletons slot straight into a ScrollTrigger workflow
• brandkit + imagegen-frontend-web for client-facing mockups and brand boards before writing code
• redesign-skill as a protocol for paid "refresh this site" jobs

Security

• Prompt-only: the only scripts are a trivial path-echo skill.sh and README asset-resizing helpers that never run at install/use time.
• Scanned clean: no injection directives, no zero-width/hidden characters, no network calls, no data exfiltration. The only "installs" it ever suggests are mainstream npm design systems (shadcn, Fluent, Carbon).

Verdict: safe and genuinely good — one of the two repos on this list you can install today with no precautions. Best for marketing sites and portfolios, not data-heavy dashboards.

What it is

Microsoft's (AutoGen team) utility that converts almost any file — PDF, Word, Excel, PowerPoint, images, audio, HTML, ZIP, YouTube URLs — into structured Markdown optimized for LLM consumption. Preserves headings, tables and links; prioritizes token-efficiency over pixel fidelity. The de facto standard doc-to-Markdown tool.

What's inside

• packages/markitdown — core library + CLI with per-format converters (pdfminer, mammoth, openpyxl, python-pptx…)
• packages/markitdown-mcp — MCP server exposing convert_to_markdown(uri) to Claude Desktop/Code
• packages/markitdown-ocr — LLM-vision OCR plugin for images embedded in documents
• Optional Azure Document Intelligence integration and LLM image captioning

Install & use

pip install 'markitdown[all]'
markitdown client-brief.pdf -o brief.md
pip install markitdown-mcp && markitdown-mcp     # give Claude "read any document"

How you can use it

• RAG ingestion: normalize client PDFs/Office docs to Markdown before chunking/embedding
• Feed briefs, spreadsheets and decks into Claude context without manual copy-paste
• Bulk-migrate legacy docs (Outlook MSG, old XLS, EPub) into Markdown knowledge bases

Security

• Clean codebase: no telemetry, no eval/exec, XML parsed via defusedxml (XXE mitigated), plugins disabled by default, known-CVE ExifTool versions refused.
• Caution (server use only): convert() happily fetches http:/file: URIs with no filtering — wrapping it in an internet-facing service creates an SSRF/local-file-read surface. Use convert_local()/convert_stream() and gate inputs. ZIP handling has no bomb protection.

Verdict: adopt without hesitation for local/CLI use; just never expose raw convert() to untrusted input on a server.

What it is

An installer/health-checker/router that gives AI agents read & search access to 13+ platforms — Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, LinkedIn, RSS — with "zero API fees." It doesn't scrape anything itself: it installs upstream open-source CLIs (yt-dlp, twitter-cli, gh…) and teaches your agent to call them directly.

What's inside

• cli.py (1,834 lines) — install/doctor/configure; auto-installs gh, Node.js, yt-dlp, twitter-cli, bili-cli, rdt-cli
• channels/ — per-platform probes with ordered fallback backends; generic web reads via Jina Reader
• cookie_extract.py — pulls Twitter/XHS/Bilibili cookies from your local browser to authenticate scraping
• skill/SKILL.md — instructions installed into your agent's skills directory

Install & use

pipx install https://github.com/Panniantong/agent-reach/archive/main.zip
agent-reach install --safe --dry-run     # see what it would change first
agent-reach install --channels=reddit,youtube
agent-reach doctor

How you can use it

• Social listening: brand/product sentiment across Twitter, Reddit, XiaoHongShu from one prompt
• Content research: YouTube/Bilibili transcript summaries, podcast-to-text, RSS monitoring
• China-market research (XHS reviews, Bilibili, Xueqiu) without Chinese-app friction

Security

• No malware or telemetry found; credentials stored locally with 0600 permissions; update checks hit GitHub only.
• Caution: heavy third-party trust delegation — every generic URL you read routes through Jina's r.jina.ai, all web searches go to Exa's hosted MCP, podcast audio uploads to Groq. Those services see your queries.
• Caution: the installer executes NodeSource's setup shell script (curl-then-bash as root, in effect), installs itself from an unpinned main.zip, and the recommended flow — "paste this URL to your agent" — has your agent fetch and obey remote instructions. Whoever controls that repo controls future installs.
• Caution: cookie-based scraping of X/Reddit/Meta violates their ToS; the README itself advises burner accounts. Cookies can also transit an optional third-party residential proxy.

Verdict: useful research glue, but run it with --safe, burner accounts, and ideally in a sandbox/VM — never with your main-account cookies.

What it is

A privacy-first research assistant you run yourself: collect sources (PDFs, web pages, video/audio), chat with them, generate notes and multi-speaker podcasts. Works with 18+ AI providers (OpenAI, Anthropic, Ollama, LM Studio…) so you're not locked to Google — or to the cloud at all.

What's inside

• FastAPI REST API (port 5055, ~24 routers) + Next.js web UI (port 8502) + SurrealDB (port 8000)
• Podcast engine: 1–4 custom speakers with episode/speaker profiles
• "Transformations" — AI content pipelines (summaries, insights) + background worker for embeddings
• Optional Docling OCR and Crawl4AI scraping for ingestion

Install & use

curl -O https://raw.githubusercontent.com/lfnovo/open-notebook/main/docker-compose.yml
# EDIT: set OPEN_NOTEBOOK_ENCRYPTION_KEY + OPEN_NOTEBOOK_PASSWORD
docker compose up -d      # UI at http://localhost:8502, API docs at :5055/docs

How you can use it

• Private research vault per project — chat and search across dumped docs/videos/articles
• Per-client knowledge bases with REST-API automation into your own tooling
• Turn research into multi-speaker podcast episodes; run fully local via Ollama for zero API spend

Security

• Good bones: provider API keys encrypted at rest (Fernet), no telemetry, SSRF guard blocks cloud-metadata IPs, private vulnerability reporting policy.
• Caution: the web UI and API are unauthenticated by default — auth only activates if you set OPEN_NOTEBOOK_PASSWORD. Compose ships SurrealDB as root:root and a placeholder encryption key. CORS is wide open. Fine on localhost; dangerous if you port-forward or host it unhardened.

Verdict: the best self-hosted NotebookLM today. Run it on localhost or behind a reverse proxy with a real password and DB credentials.

What it is

Turns an AI coding CLI (Claude Code, Codex, OpenCode…) into a job-search command center: scans 150+ company ATS portals for new listings, has the LLM score each job against your CV (1.0–5.0 with an A–F rubric), generates tailored ATS-optimized CV/cover-letter PDFs, and tracks your pipeline in local Markdown files. Built by a developer after a real 740-listing job hunt.

What's inside

• modes/*.md — ~30 prompt "modes" (auto-pipeline, apply, deep research, interview prep) in 15+ languages
• scan.mjs + 61 provider modules — zero-token HTTP scanner for Greenhouse, Lever, Ashby, Workday…
• Playwright layer for read-only job-description extraction and HTML→PDF resume generation
• Go terminal dashboard + alpha web view — both read only local files; Notion/Gmail plugins

Install & use

npx @santifer/career-ops init
cd career-ops && claude              # chat onboarding builds cv.md + profile
/career-ops scan                     # find new listings
/career-ops <paste job URL>          # score → report → tailored PDF → tracker

How you can use it

• Serious job hunting: filter hundreds of listings, apply only to 4.0+/5.0 fits
• Reverse it for recruiting: score candidate–role fit at scale
• Re-aim the clean ATS-API provider library at competitor hiring pages for market/lead intel

Security

• Local-first done right: CV/profile/data gitignored, CI blocks personal data in PRs, no telemetry, SSRF guard on URL checks, Playwright layer marked strictly read-only.
• Caution: "never auto-submit" is enforced only in prompts, not code — models can misbehave; keep a human on every submit. Optional eval scripts send your full CV to Gemini/OpenAI/OpenRouter. Portal scraping uses a browser-impersonating user agent — ToS risk is explicitly yours. Self-updater runs npm install (supply-chain surface).

Verdict: well-engineered and genuinely useful for a high-volume search — with human review before every application.

What it is

"PM Skills Marketplace" by Paweł Huryn (productcompass.pm): 68 skills and 42 slash-command workflows across 9 Claude Code/Cowork plugins. Each skill is a prompt template encoding a named product-management framework (Teresa Torres, Cagan, JTBD, RICE, SWOT…) that walks the AI through producing a structured deliverable. No application code ships with it.

What's inside

• pm-product-discovery (13): ideation, assumption mapping, experiments, interview synthesis
• pm-product-strategy (12): vision, lean canvas, SWOT, PESTLE, Porter's Five Forces, pricing
• pm-execution (16): PRDs, OKRs, roadmaps, sprint plans, pre-mortems, retros
• pm-market-research (7), pm-go-to-market (6), pm-marketing-growth (5), pm-data-analytics (3)
• pm-ai-shipping: static security/performance audit commands for AI-built code; pm-toolkit: NDAs, privacy policies

Install & use

claude plugin marketplace add phuryn/pm-skills
claude plugin install pm-execution@pm-skills     # install per plugin
/write-prd    /discover    /strategy    /plan-launch

How you can use it

• Client-ready PRDs, user stories and test scenarios from a rough brief (/write-prd, /red-team-prd)
• Prioritize feature backlogs with RICE/ICE instead of gut feel
• Package strategy artifacts (lean canvas, GTM plan, battlecards, pricing) as billable client deliverables

Security

• 147 files, zero executable content in the installed plugins — no hooks, no scripts, no binaries, no MCP servers. No prompt-injection directives found; nothing writes outside the project.

Verdict: safe to install today. Best for solo devs/agencies wearing the PM hat; pm-execution + pm-product-discovery + pm-go-to-market cover most needs.