Today's drop: 1 new tool, 3 listing updates
The directory lists 150 hand-checked tools. New today: claude-code-hooks, a free, open-source set of Claude Code plugins that check every shell command, file read and edit before it runs and block the destructive or secret-leaking ones. Updated: Qwen Code (v0.27.4, safer Auto Mode), CISO Assistant (v4.3.1) and CVEngine (v1.4). AI for IT & Cybersecurity grows to 63 tools; we retired one Research listing to stay at 150. Pricing only where the official page states it.
Previous drop: Nov 14, 2026 →
Get the weekly dropThe best new AI tools, one email a week. Free.
New in the directory 1
Browse all 150Tool
Category
Price
Status
Link
claude-code-hooksFree, open-source (MIT) collection of ready-made hooks for Claude Code from Karan Bansal, Head of AI at the application security company ArmorCode, shipped as a 22-plugin marketplace you install from inside Claude Code (/plugin marketplace add karanb192/claude-code-hooks, then for example /plugin install block-dangerous-commands@claude-code-hooks). The safety plugins run as separate processes before a tool call executes, so the allow-or-deny decision is made by ordinary code rather than the model: block-dangerous-commands stops commands such as rm -rf ~, fork bombs, curl | sh and force pushes to main; protect-secrets blocks reading, editing or exfiltrating .env and other credential files; git-safety adds branch-aware git guardrails; protect-tests stops the agent deleting or skipping tests to fake a green build; config-guard stops it tampering with its own hook and settings files; instructions-audit locks the session when a loaded CLAUDE.md or rules file carries hidden directives. Other plugins handle session logs, auto-formatting, Slack alerts and context-cost tracking. Every hook is mapped to the OWASP LLM Top 10 2026, and the repo reports 1,949 passing tests. Pattern-based rules can be bypassed (the author says he found two bypasses of his own secrets check) and hooks cannot stop a malicious session-start hook already on disk, so use it alongside scoped credentials and sandboxing, not instead of them. Claude Code only. About 580 GitHub stars; ArmorCode published a write-up on Oct 29, 2026, press coverage followed on Nov 14, and the author plans a Black Hat Europe Arsenal demo on Dec 9-10.
Price Free and open source (MIT)
Why this one
- claude-code-hooks puts the permission check outside the model. Coding agents rarely get hacked in the stories people share; they pick a destructive step on their own, like an
rm -rfthat ends in~/or agit reset --hardover a night of uncommitted work. Claude Code's hooks run a separate program before each tool call, and this repo is a tested set of those programs:block-dangerous-commands(rm -rf ~, fork bombs,curl | sh, force pushes to main),protect-secrets(.env files and other credentials),git-safety,protect-tests(no deleting or skipping tests to fake a green build),config-guard(the agent cannot edit its own guardrails) andinstructions-audit(hidden directives in CLAUDE.md files). The decision is plain code, so the same command gets the same answer every time and a prompt injection cannot argue its way past it. It is by Karan Bansal, Head of AI at ArmorCode, MIT licensed, with about 580 GitHub stars, 22 installable plugins and 1,949 passing tests mapped to the OWASP LLM Top 10 2026. It is not new this week (first released in January), but it got fresh coverage on Nov 14 and it is the easiest way we have seen to give Claude Code sane guardrails in two commands. Be honest about the limits: pattern rules can be bypassed (the author found two bypasses of his own secrets check), hooks cannot stop a malicious session-start hook already on disk, and it only covers Claude Code; for other agents see HOL Guard or Failproof AI.
Listing updates
Checked on Nov 15, 2026 against official pages and release notes.
- Qwen Code: v0.27.4 (Nov 15) makes Auto Mode ask for explicit authorization before risky actions, reviews classifier timeouts and visibly marks Full Access. It also restores the
/memorycontrols in the CLI, lets managed agent sessions use local folders you grant without saving credentials or paths, adds Feishu and WeCom webhook output and patches vulnerable dependencies (MCP, DOMPurify and compressed-request parsing). The 1.0 line is still a preview. - CISO Assistant: v4.3.1 (Nov 14) makes the optional AI chat index evidence attachments and published documents, so it can answer from them; fixes MCP sign-in failing when an app requests scopes it may not have; adds an outline of ISO 9001:2015 for quality management (English and French); and improves how tasks are shown.
- CVEngine: v1.4.0 and v1.4.1 (Nov 14) make the status page, project rows and headline counts much faster on small servers, draw each registered domain as its own cluster on the graph and skip restarts when an update only changes documentation. About 730 GitHub stars now, up from about 400 when we listed it in October.
Category changes
- AI for IT & Cybersecurity grows from 62 to 63 tools with claude-code-hooks.
- Productivity & Research drops from 9 to 8 tools: to stay at the 150-tool cap we retired Consensus. Elicit covers AI search and analysis of research papers for our readers, and we never had verified pricing for Consensus. Its old page redirects to the Research category.
Checked, not listed
- Vaikora Guard MCP (Data443): the MCP policy proxy with a hash-chained audit log, now MIT on PyPI (Data443 blog, Nov 14); still watching for traction.
- bastyn: a new Apache-2.0 static scanner for MCP configs and agent instructions; alpha, with anonymous telemetry on by default (turn it off with
--offline). - Bright Lines: a reported commit-approval hook for Claude Code; we could not confirm the repository yet.
- sapper-v2: an autonomous AI pentest project with big competition claims we could not verify.
- Routine: Codex 0.171 alphas, Orca v1.4.259, Hermes Agent canaries; no new Claude Code release since v2.1.325. Ollama v0.40.17 and Wazuh 4.14.11 are still release candidates.