Today's drop: 3 new tools, 2 listing updates
The directory now lists 112 hand-checked tools. New today: NVIDIA's SkillSpector, a free scanner that checks AI agent skills before you install them; Exa, the web search API many AI agents use, now on Exa 3.0 with a $4-per-1,000 Lite tier; and MiniMax M3.1 Flash, a big open-weight agent and coding model. Updated: Claude (Claude Duet) and Kling AI (Kling 4.0). AI for IT & Cybersecurity grows to 32 tools. Pricing only where the official page states it.
Previous drop: Oct 23, 2026 →
Get the weekly dropThe best new AI tools, one email a week. Free.
New in the directory 3
Browse all 112Tool
Category
Price
Status
Link
SkillSpector (NVIDIA)NVIDIA's open-source scanner that checks AI agent skills (for Claude Code, Codex, Gemini CLI and MCP) before you install them. It looks for 71 vulnerability patterns across 17 categories, including prompt injection, data exfiltration, privilege escalation, supply-chain risk and MCP tool poisoning, then gives a 0-100 risk score with JSON, Markdown or SARIF reports. Fast static checks work offline; an optional LLM pass adds deeper review. Scans Git repos, URLs, zip files or folders, and runs in Docker. About 21,000 GitHub stars; v2.13.1 shipped Oct 23, 2026.
Price Open source (Apache-2.0); optional LLM analysis billed by your AI provider
ExaWeb search API built for AI apps and agents: runs its own crawler and index, returns ranked results with clean page text or short highlights, and adds deep research, monitors and an agent API. Exa 3.0 (Oct 19, 2026) brings a larger embedding model, generally available Dynamic Highlights and billions of new documents such as products, local businesses and events, plus a cheaper Lite search type. Exa says it ranks #1 on the Artificial Analysis Search Index. Also available as a hosted MCP server.
Price Pay-as-you-go: Lite search $4 per 1,000 requests, Auto $7 per 1,000; free tier $10 in credits each month
MiniMax M3.1 FlashMiniMax's open-weight multimodal mixture-of-experts model (about 428B total, 23B active per token, up to 1M tokens of context) for coding, agentic work, computer use and cybersecurity tasks. Takes text, image and video input. MiniMax's own benchmarks put it at 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0, close to big closed models on several agent tests. Too big for a home PC: MiniMax recommends serving it with SGLang, vLLM or TokenSpeed, or using it through MiniMax Agent and the MiniMax API. New in October 2026.
Price Free open weights (MiniMax Community License: commercial use needs a 'Built with MiniMax M3.1-Flash' credit and a notice to MiniMax; written approval above $20M yearly revenue)
Why these three
- SkillSpector answers one question: is this agent skill safe to install? Skills for Claude Code, Codex, Gemini CLI and MCP run with a lot of trust, and NVIDIA's README says that in the 31,132 skills it analyzed for its research, 26.1% had vulnerabilities and 5.2% showed likely malicious intent. SkillSpector checks 71 patterns across 17 categories (prompt injection, data exfiltration, privilege escalation, supply chain, MCP tool poisoning and more) and gives a 0-100 risk score, with SARIF output for CI. The static scan (
--no-llm) runs offline; an optional LLM pass goes deeper. Apache-2.0, about 21,000 GitHub stars, v2.13.1 shipped Oct 23. It's part of NVIDIA's Verified Skills pipeline. - Exa is a search API built for AI apps: it runs its own crawler and index instead of wrapping Google or Bing, and returns results with clean page text or short highlights an agent can read. Exa 3.0 (Oct 19) adds a larger embedding model, makes Dynamic Highlights generally available (Exa says 95% fewer tokens than full page text) and adds billions of documents such as products, local businesses and events. The new Lite search costs $4 per 1,000 requests, Auto is $7, and the free tier gives you $10 in credits every month. Exa says it ranks #1 on the Artificial Analysis Search Index; its other benchmark wins are its own.
- MiniMax M3.1 Flash is MiniMax's new open-weight model for coding, agents, computer use and security work: about 428B parameters in total with 23B active, up to 1M tokens of context, and text, image and video input. On MiniMax's own benchmark table it scores 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0, close to big closed models on several agent tests but behind them on others. It's a data-center model (SGLang, vLLM or TokenSpeed), not a laptop one. The license is free for non-commercial use; commercial use needs a "Built with MiniMax M3.1-Flash" credit and a notice to MiniMax, with written approval above $20M in yearly revenue.
Listing updates
Checked on Oct 24, 2026 against official pages and announcements.
- Claude: Anthropic opened a research preview of Claude Duet, a separate proactive agent app (iOS, Android and claude.ai). It sends a morning brief from your email and calendar, works on tasks on its own cloud computer with its own browser, and shows you what it's about to do before it sends a message, makes a payment, signs a document or deletes anything. Waitlist only, for US Pro and Max plans (not Team or Enterprise yet), and its background checks count toward your plan's usage limits.
- Kling AI: Kling 4.0 is fully available on kling.ai (since Oct 15). It generates up to 30 seconds in one pass, handles fast motion and complex camera moves more steadily, and Omni Reference combines up to 15 reference assets and up to 10 keyframes. It also works through Kling MCP and CLI; 10-bit HDR output and developer API access are listed as coming soon.
Category changes
- AI for IT & Cybersecurity grows from 31 to 32 tools. The Breaking into IT guide now suggests scanning any skill or MCP server with SkillSpector before you install it for your AI coding agent.
- Productivity & Research adds Exa, and Coding adds MiniMax M3.1 Flash next to Muse Spark Open as another large open-weight model.
Checked, not listed
- Cisco skill-scanner and Snyk agent-scan: both are solid scanners for agent skills and MCP servers (about 2,600 and 3,200 GitHub stars). We picked SkillSpector as the first one to list because it's the most widely used and has a clear Apache-2.0 license; Cisco's repo doesn't show a standard license on GitHub.
- AuroraTTS-100K: an English text-to-speech model with under 100,000 parameters (127KB at INT8) that runs many times faster than real time on a CPU. A fun research release, but it's new and small, may slur hard words, and Pocket TTS already covers CPU text-to-speech here.
- Ling-3.0-flash: the base model came out in August; this week brought fine-tunes, not a new release.