moshi-hooks

Claude Code hook adapter that bridges agent lifecycle events to the Moshi API, which fans out push notifications to update the Live Activity on your iPhone.

Claude Code → hooks (stdin) → moshi-hooks → Moshi API → APNs → Live Activity

Install

Requires Bun (Node.js support planned). Zero runtime dependencies.

bun i -g moshi-hooks

Or run directly without installing:

bunx moshi-hooks setup

Setup

moshi-hooks token <YOUR_TOKEN>
moshi-hooks setup              # user scope (~/.claude/settings.json)
moshi-hooks setup .            # project scope (.claude/settings.json in cwd)
moshi-hooks setup /path/to/dir # project scope (absolute path)

Project-scoped hooks can be committed to the repo and shared with your team. Both scopes are idempotent — safe to run multiple times. Existing hooks from other tools are preserved.

To remove:

moshi-hooks uninstall              # user scope
moshi-hooks uninstall .            # project scope
moshi-hooks uninstall /path/to/dir # project scope

All hooks run with async: true so they never block Claude.

Event mapping

| Hook Event | eventType | category | Sends to API? | |---|---|---|---| | SessionStart | — | — | No (persists model to state file) | | Stop | stop | task_complete | Yes (visible push) | | SubagentStop | agent_turn_complete | info | Yes (visible push) | | Notification | notification | approval_required | Yes (visible push) | | PreToolUse | pre_tool | tool_running | Yes, filtered (silent) | | PostToolUse | post_tool | tool_finished | Yes, filtered (silent) | | UserPromptSubmit | — | — | No (skipped) |

Tool events are filtered to only fire for: Bash, Edit, Write, WebFetch, WebSearch, Task. Read-only tools (Read, Glob, Grep) are skipped.

How it works

Each hook invocation is a separate process. Claude Code pipes JSON to stdin, the adapter reads it, maps the event, and POSTs to the Moshi API.

Cross-event state is persisted to /tmp/moshi-hook-{session_id}.json so that later events (like Stop) can include the model name and last tool from earlier events (like SessionStart and PreToolUse).

Context window usage is estimated by reading the last ~10KB of the transcript JSONL and parsing the most recent usage data.

Testing

bun test

Typecheck

bun typecheck