Today's drop: 2 new tools, 3 listing updates
The directory lists 150 hand-checked tools. New today: safeagents, a free, open-source proxy from Hugging Face that checks every tool call your coding agent makes and blocks the risky ones before the agent runs them, and Yogurt, free, open-source AI meeting notes that are recorded, transcribed and summarized on your own computer. Updated: CISO Assistant (v4.3.2), Failproof AI (v1.0.14) and SkillSpector (2.13.5). AI for IT & Cybersecurity grows to 64 tools; we retired two listings to stay at 150. Pricing only where the official page states it.
Previous drop: Nov 15, 2026 →
Get the weekly dropThe best new AI tools, one email a week. Free.
New in the directory 2
Browse all 150Tool
Category
Price
Status
Link
safeagents (Hugging Face)Free, open-source (Apache-2.0) local proxy from Hugging Face that sits between your AI coding agent and its model server and checks every tool call the model asks for before the agent sees it. A call is blocked if it uses a tool the request never declared, reaches a host that appears nowhere in the conversation, or trips one of Hugging Face's atif-scan detectors (for example reading /proc/1/environ or a Kubernetes service-account token, curl --noproxy or nsenter); a blocked call is swapped for a text notice, so the agent runs nothing. Borderline calls can go to an optional judge model that sees only what you wrote, the call and the findings, and can add blocks but never lift one; Hugging Face publishes a benchmark of 661 real flagged tool calls to help you pick one. Works with any agent that lets you set a base URL over OpenAI Chat Completions or Anthropic Messages: setup guides cover Claude Code, OpenCode, Qwen Code, Cline, OpenHands, Hermes Agent, OpenClaw, pi, Mistral Vibe, DeepSeek Harness and GitHub Copilot in VS Code (bring-your-own-key models only). Codex CLI (Responses API) and Gemini CLI calls are not checked yet. Start with --log-only to see what it would block before you turn blocking on. A separate scrub command masks credentials, emails, home directories and public IP addresses in agent session logs before you share them. Installs with a one-line script (it uses uv) on macOS, Linux, WSL or Windows PowerShell. About 286 GitHub stars (repo created Oct 28, 2026); no tagged releases yet, with commits as recent as Nov 15.
Price Free and open source (Apache-2.0); an optional judge model is billed by your AI provider
YogurtFree, open-source (MIT) desktop app for private AI meeting notes from Frigade (backed by Y Combinator), pitched as an alternative to Granola. It records your microphone and the call audio playing on your computer, so no bot joins the meeting, transcribes as you talk with whisper.cpp (plus Apple SpeechAnalyzer on supported Macs; 99 languages in multi-language mode) and writes the summary with a local open-weight model through llama.cpp. Audio, transcripts and notes stay on your computer, and it works offline after a one-time model download (about 5 GB first, then 14 GB or more for the best model your machine can run). Ask questions across meetings with links to the source, search by folder, title or transcript, export Markdown, and let an assistant app read your notes through a local MCP server when you allow it. No account needed. Runs on Mac (Apple silicon, macOS 14.4+), Windows 10 (2004) or newer and Linux (Ubuntu 22.04+ AppImage) with at least 8 GB of memory; calendar and speaker names from Zoom, Teams and Meet are Mac only for now, and the Windows build is not yet code-signed. Public since Nov 10, 2026, with about 1,760 GitHub stars in under a week and frequent releases (v0.3.31 on Nov 16). Let everyone on the call know you are recording.
Price Free and open source (MIT); no account needed
Why these two
- safeagents checks tool calls at the one point every agent shares: the connection to the model. It runs as a small local proxy, you point your agent's base URL at it, and each tool call the model returns is checked before the agent ever sees it. A call is blocked if it uses a tool the request never declared, if it reaches a host that appears nowhere in the conversation (a common sign of data being sent out), or if Hugging Face's atif-scan detectors flag it, such as reading
/proc/1/environor a Kubernetes service-account token. Blocked calls are replaced with a text notice, so nothing runs. That makes it agent-neutral: the setup guides cover Claude Code, OpenCode, Qwen Code, Cline, OpenHands, Hermes Agent and more, while hook-based guards like claude-code-hooks work inside one agent. Start with--log-onlyto see what it would block. The optional judge model only sees your request, the call and the findings, never the agent's own reasoning, so a hijacked agent cannot argue its case. We passed on it in early November at about 70 stars; it is now near 290, with plugins loadable from the Hugging Face Hub and a public benchmark for judge models. Limits: Codex CLI and Gemini CLI calls are not checked yet, there are no tagged releases, and the installer is a script piped to your shell, so read it first. - Yogurt is the meeting notetaker for people who cannot send call audio to someone else's cloud. It records your microphone and the call audio on your own machine (no bot joins), transcribes with whisper.cpp and summarizes with a local model through llama.cpp, so recordings, transcripts and notes never leave the computer, and it works offline after the model download. It is free, MIT-licensed and needs no account, from Frigade (backed by Y Combinator), and it passed about 1,760 GitHub stars within a week of going public on Nov 10. For an IT team or MSP it is an easy answer to "can I use an AI notetaker on client calls?". Plan for the download (about 5 GB, more for the best model) and 8 GB of RAM; speaker names and calendar features are Mac only for now, and the Windows build is not code-signed yet. Always tell people on the call you are recording.
Listing updates
Checked on Nov 16, 2026 against official pages and release notes.
- CISO Assistant: v4.3.2 (Nov 15) adds MCP tools that let an AI assistant read, create and update documents in the GRC platform, and fixes missing audit observations in Word exports, metric quick input ignoring the date and unclear OIDC consent messages.
- Failproof AI: v1.0.14 (Nov 15) stops losing telemetry during outages. Event batches are now kept on disk and resent once the network or server is back, instead of being marked failed after three tries; batches the server rejects for good are kept for inspection. About 9,100 GitHub stars.
- SkillSpector (NVIDIA): 2.13.5 (Nov 16) is a maintenance release that improves bounded parsing of large agent skills and structured data and keeps security findings across decoded views. NVIDIA repeats that a warning or incomplete scan is not proof a skill is safe. About 26,300 GitHub stars.
Category changes
- AI for IT & Cybersecurity grows from 63 to 64 tools with safeagents.
- Productivity & Research grows from 8 to 9 tools with Yogurt.
- Image drops from 8 to 7 tools: to stay at the 150-tool cap we retired Recraft. Ideogram, Adobe Firefly and Midjourney cover what our readers use it for, and we never had verified pricing for it. Its old page redirects to the category.
- Writing drops from 4 to 3 tools: we also retired QuillBot. Grammarly covers grammar checks and rewriting, and we never had verified pricing for QuillBot. Its old page redirects to the category.
Checked, not listed
- Vigil SOC 1.1 (DeepTempo, Nov 11): an Apache-2.0 AI SOC with no-code agent manifests; about 440 GitHub stars, and AiSOC already covers open-source AI SOC triage for most readers.
- Vaikora Guard MCP (Data443): a second blog post (Nov 15) on the MIT release; still watching for traction.
- Routine: no new Claude Code release since v2.1.325; Codex 0.171 alphas; Qwen Code nightlies; CVE Intel v0.3.37 and v0.3.38 (README and demo videos only); n8n 1.123.97 (1.x maintenance line). Ollama v0.40.17 and Wazuh 4.14.11 are still release candidates.