AiOps Community hooks and monitors

Things that run automatically in your Claude Code session, packaged as plugins so they can be installed with one command and updated later.

Two kinds live here, and they are mirror images of each other. A hook reacts to something Claude does: it fires on a lifecycle event and can block, format or record. A monitor reacts to something the world does: it runs in the background for the whole session and every line it prints reaches Claude as a notification, so Claude learns about a failing build or a log entry without anyone asking it to look.

Claude Code only. Unlike skills, there is no file to upload in claude.ai. Both are configuration that Claude Code executes, so the only shareable form is a plugin containing a hooks/hooks.json or a monitors/monitors.json. That is what every entry here is.

They share a repository because they share a trust model: both run unsandboxed, automatically, with your credentials. That is what CONTRIBUTING.md is written around. Skills and agents, which only influence what Claude says, live in AiOpsCommunity/skills.

Hooks

Hook	Event	Description
block-secret-files	PreToolUse	Blocks reads and writes of secret files (.env, private keys, credential stores) before the tool runs.
claude-md-maintainer	Stop	Collects candidate CLAUDE.md lines into a local, git-excluded inbox; a bundled skill verifies them, prunes stale lines and enforces a size budget.
git-guardrails	PreToolUse	Blocks irreversible git operations: force pushes to protected branches, hard resets over uncommitted work, protected branch deletion, untracked file wipes and history rewrites.
project-checks	PostToolUse	Runs the formatter or linter a project already defines on the file just edited, returning failures so they are fixed in the same turn. Runs only in directories you mark as trusted.

Monitors

Monitor	Watches	Description
ci-watch	GitHub Actions	Reports a finished workflow run for the current branch. Failures always; successes only when it saw the run start.
log-watch	Log files	Reports only the lines matching patterns you configure, rate limited so a noisy log cannot flood the session.

Install

/plugin marketplace add AiOpsCommunity/hooks
/plugin install <hook-name>@aiops-hooks
/reload-plugins

Verify with /hooks, which lists every registered hook and where it came from. Update later with /plugin marketplace update aiops-hooks.

Without the marketplace

Copy the scripts somewhere stable and register the hook yourself in ~/.claude/settings.json (personal) or .claude/settings.json (one project). Each hook's own README shows the settings block, since ${CLAUDE_PLUGIN_ROOT} does not exist outside a plugin and has to be replaced with a real path.

Repo layout

.
├── README.md
├── CONTRIBUTING.md
├── .claude-plugin/
│   └── marketplace.json          # plugin catalog for /plugin install
├── .github/
│   └── workflows/
│       └── pr-checks.yml         # runs the scripts below on every PR
├── scripts/
│   ├── validate.sh               # structure: JSON, events, paths, permissions
│   ├── pr-policy.sh              # one hook per PR, registered properly
│   └── review-flags.sh           # advisory: lines a reviewer should read
├── hooks/
│   ├── _template/                # copy this to start a new hook
│   └── <hook-name>/
│       ├── .claude-plugin/
│       │   └── plugin.json       # manifest, metadata only
│       ├── hooks/
│       │   └── hooks.json        # the hook config (same shape as settings.json)
│       ├── scripts/
│       │   └── <script>          # what the hook actually runs
│       ├── skills/               # optional, only when it serves the hook
│       │   └── <name>/SKILL.md
│       └── README.md             # what it does, why, how to test, how to disable
└── monitors/
    ├── _template/                # copy this to start a new monitor
    └── <monitor-name>/
        ├── .claude-plugin/
        │   └── plugin.json
        ├── monitors/
        │   └── monitors.json     # array of monitor entries
        ├── scripts/
        │   └── <script>          # the long-running command
        └── README.md

The nesting looks odd at first: hooks/<hook-name>/hooks/hooks.json. The outer directory is this repo's catalog. The inner one is where Claude Code looks inside a plugin — hooks/hooks.json and monitors/monitors.json are both defaults it discovers on its own. Manifests live in .claude-plugin/, everything functional lives at the plugin root.

Add a new hook

1. Copy the template and rename it (kebab-case, this becomes the plugin name):

   cp -r hooks/_template hooks/my-new-hook

2. Fill in .claude-plugin/plugin.json, hooks/hooks.json, the script, and the README.

3. Add a plugin entry to .claude-plugin/marketplace.json and a row to the table above.

4. Run ./scripts/validate.sh and test locally (see CONTRIBUTING.md).

Add a new monitor

Same shape, different directory and config file:

cp -r monitors/_template monitors/my-new-monitor

Fill in .claude-plugin/plugin.json and monitors/monitors.json, write the script, register it in marketplace.json and the Monitors table, and validate. The template README explains the contract: stdout is a message to Claude, so print only what deserves to interrupt.

Monitors run only in interactive CLI sessions, run unsandboxed at the same trust level as hooks, and are an experimental component whose schema may still change. Note that ${user_config.*} is not substituted into a monitor command and monitor processes do not receive CLAUDE_PLUGIN_OPTION_<KEY>, so a monitor that needs configuration reads it from a file the script owns.

Events you can hook into

Event	Fires	Typical use
PreToolUse	Before a tool runs, can block it	Guardrails, policy checks
PostToolUse	After a tool succeeds	Formatting, linting, tests
UserPromptSubmit	When a prompt is submitted	Inject context, reject prompts
Notification	On a Claude Code notification	Desktop alerts, Slack pings
Stop	When the main agent finishes	Wrap-up work, changelog updates
SubagentStop	When a subagent finishes	Collect subagent output
SessionStart	On start or resume	Load branch state, env checks
SessionEnd	On session end	Cleanup, logging
PreCompact	Before context compaction	Persist state that would be lost

That table is the common set, not the whole list — Claude Code ships around thirty events, including PostToolUseFailure, PermissionRequest, FileChanged, SubagentStart and TaskCompleted. validate.sh only warns about an event it does not recognise, so a hook on a newer event is never blocked by this repo being a release behind.

matcher filters within an event. For tool events it matches the tool name and accepts regex alternation (Write|Edit); MCP tools are matched as mcp__<server>__<tool>. Leave it out to match everything.

Things worth knowing

Feature	What it gives you
if: "Bash(git *)"	Narrows a tool hook further than matcher can, using permission-rule syntax. Strips NAME=value prefixes and fails open when it cannot parse.
args (exec form)	Spawns the script directly instead of through sh -c. No quoting traps. See CONTRIBUTING.md.
userConfig	Values Claude Code prompts for at enable time, stored in user settings and passed to the hook as CLAUDE_PLUGIN_OPTION_<KEY>. The right place for anything that decides what a hook may do.
defaultEnabled: false	Ships the plugin installed but switched off, for hooks that add cost or scope.
async / asyncRewake	Runs a slow hook in the background; asyncRewake wakes Claude on exit 2 with the output.
${CLAUDE_PLUGIN_DATA}	A per-plugin directory that survives updates, unlike ${CLAUDE_PLUGIN_ROOT}.

Check the hooks reference and the plugins reference before relying on details here; those pages are the source of truth and this is a summary.