tool-time

tool-time

Tool usage analytics for Claude Code, Codex CLI, and OpenClaw.

What this does

Hooks capture every tool call in your Claude Code sessions. Transcript parsers extract tool usage from Codex CLI and OpenClaw sessions. A Python script aggregates 7 days of stats. An AI agent reads those stats, spots patterns (high error rates, edit-without-read, Bash overuse), checks your CLAUDE.md for gaps, and proposes concrete fixes.

Optionally, share anonymized stats to a community dashboard at tool-time.org and compare your patterns against other users.

Installation

First, add the interagency marketplace (one-time setup):

/plugin marketplace add mistakeknot/interagency-marketplace

Then install the plugin:

/plugin install tool-time

Usage

/tool-time

The agent regenerates your stats, analyzes tool usage patterns for the current project, flags issues, reads your CLAUDE.md/AGENTS.md, and proposes specific edits. It also recommends relevant skills from playbooks.com based on your project language.

/tool-time rig

Rig audit mode: joins your usage data against every registered MCP server and enabled plugin (global config, project config, plugin cache), plus a live process snapshot — then recommends what to disable. Surfaces zero-use servers with their RAM cost, duplicate registrations (the same server reachable two ways), and idle plugins. rig.py collects the data; the agent makes the recommendations and asks before touching any config.

What gets flagged

• High error rates: tools failing 10%+ of the time
• High rejection rates: tools the user frequently denies
• Edit-without-read: editing files that weren't read first (a surprisingly common source of bugs)
• Bash dominance: Bash >50% of calls, but only when Bash is doing file reads (cat, grep, find). Git, test runners, and deployment scripts are legitimate Bash usage.
• Low tool diversity: underutilizing available tools
• Rig waste (rig mode): zero-use MCP servers still spawning processes per session, duplicate registrations, idle plugins

Proactive digest

You don't have to remember to run /tool-time. On SessionEnd the pipeline refreshes stats and writes a one-to-two-line digest (maintain.py); on SessionStart the hook surfaces it as session context — but only when something is notable (pipeline broken, error-rate spike, several unused servers). Healthy rig = silent digest.

How it works

hooks (PreToolUse, PostToolUse, SessionStart, SessionEnd)
  │
  ▼
~/.claude/tool-time/events.jsonl    ← one JSONL line per event
  │
  ├──▶ summarize.py                 ← pure data aggregation (7-day window, per-project)
  │      │
  │      ▼
  │    ~/.claude/tool-time/stats.json   ← calls, errors, rejections, last_used per tool
  │      │
  │      ├──▶ skill (agent analysis)    ← reads stats + CLAUDE.md, proposes fixes
  │      │
  │      └──▶ upload.py (opt-in)        ← anonymized submission to community API
  │             │
  │             ▼
  │           Cloudflare Worker + D1    ← community dashboard
  │
  ├──▶ rig.py (on demand)           ← joins usage × MCP/plugin registrations × ps → rig.json
  │
  └──▶ maintain.py (SessionEnd)     ← log rotation, seq GC, writes digest.txt
         │
         ▼
       digest.txt ──▶ SessionStart hook surfaces it (only when notable)

summarize.py and rig.py are pure data preparation: no opinions, no thresholds. The agent does the analysis. This split keeps the data pipeline honest and the recommendations adaptable.

Known coverage limit: hooks don't fire for tool calls made inside subagents, so subagent-heavy usage undercounts in events.jsonl. The rig-audit skill cross-checks transcripts before recommending removals.

Community sharing (Opt-In)

Enable in ~/.claude/tool-time/config.json:

{ "community_sharing": true }

What's shared (strict allow-list): submission token (random hex, not your identity), timestamp (truncated to hour), per-tool call/error/rejection counts, model name, skill usage, MCP server usage, installed plugins.

What's NOT shared: file paths, project names, error messages, skill arguments.

GDPR deletion: /tool-time delete my data: looks up your token, confirms, deletes all server-side data, offers to disable future uploads.

Codex CLI + openClaw support

Codex CLI doesn't support hooks, so backfill.py parses historical session transcripts from ~/.codex/sessions/. OpenClaw transcripts are parsed from ~/.openclaw/agents/, ~/.moltbot/agents/, and ~/.clawdbot/agents/ (the app was rebranded multiple times; sessions are deduplicated across directories).

Development