Orpheus
Token adoption in teams

How do you measure Claude Code adoption across a team?

Short answer

Track five numbers each week: how many people with a seat used Claude Code, sessions per active person, each seat's share of its weekly limit used by the reset, how long agents wait on a human, and outcomes such as merged pull requests. Token totals alone say little about value and are easy to inflate.

On this page

The five metrics

Each metric answers a different question. Read them together: high sessions with flat outcomes is a different problem from low sessions with strong outcomes.

Claude Code adoption metrics, read weekly
MetricFormulaWhere it comes from
Active people per weekPeople with 1 or more sessions this week ÷ seats paidTeam/Enterprise analytics (daily active users), or a count from each person
Sessions per active personSessions this week ÷ active peopleAnalytics dashboard (sessions per day), OpenTelemetry
Share of weekly limit used at resetLast 7-day percentage before resets_at, per seat, averaged over 4 weeksStatusline rate_limits, Settings > Usage, Orpheus
Agent wait time on humansMinutes from an agent asking for input to the next promptClaude Code hooks, logged per session
OutcomesMerged PRs with Claude Code, lead time, change failure rateAnalytics contribution metrics, your Git host, DORA metrics
Sources 1 to 4 below. Checked .

Why weekly limit at reset, not tokens: on Pro, Max and Ultra, and on Team and Enterprise seats, the capacity you buy is an allowance that resets every five hours and every week,1 so the share used by the reset tells you how much of what you paid for was put to work. A seat that ends weeks at 25% is a different conversation from one that hits its limit on Thursday. The full method is in Are you using the Claude capacity you pay for?

Where each number comes from

/usage in Claude Code: one person, one machine

/usage shows plan usage bars and a breakdown by skill, subagent, plugin and MCP server. The breakdown is local.

Anthropic says

“The figures are approximate and computed from local session history on this machine, so usage from other devices or claude.ai is not included.”

Manage costs effectively, code.claude.com · Checked .

Good for coaching one person. Not a team number.

Statusline rate_limits: per login, live

A statusline script receives rate_limits.five_hour.used_percentage, rate_limits.seven_day.used_percentage and each window's resets_at in Unix seconds. The fields appear only for Pro, Max and Ultra subscribers, and only after the first API response in a session.2 This is the first-party, local way to read the weekly percentage for each account. The docs do not list Team or Enterprise seats here, so for those use Settings > Usage or the admin console.

Team and Enterprise admin consoles

Check your admin console for what your plan shows. Anthropic's docs describe a Claude Code analytics dashboard for Admins and Owners with daily active users and sessions, lines of code accepted, suggestion accept rate, and, with the GitHub app connected, pull requests with Claude Code. It also has a leaderboard of the top 10 users by contribution and a CSV export.3 Anthropic calls the contribution figures “deliberately conservative”.3 For per-user token counts and cost, the docs point to OpenTelemetry export or the spend report CSV.1

The Orpheus usage board

If you turn usage on in the connector's local policy file, Orpheus reads each login's weekly and session percentage about once an hour on the machine and shows every connected account with its weekly and 5-hour use, the forecast for the reset and the reset time. It is off by default.

Log the weekly percentage from the statusline

Without a dashboard, each person can log their own number. This script writes the latest 7-day percentage to one file per weekly window, named by its reset time, so the file left behind after the reset holds the share used by the reset (as of the last update before it).

  1. Save the script

    Save it as ~/.claude/statusline.sh and make it executable with chmod +x. It needs jq.

    #!/bin/bash input=$(cat) pct=$(echo "$input" | jq -r '.rate_limits.seven_day.used_percentage // empty') reset=$(echo "$input" | jq -r '.rate_limits.seven_day.resets_at // empty') dir="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/week-usage" if [ -n "$pct" ] && [ -n "$reset" ]; then mkdir -p "$dir" echo "$pct" > "$dir/$reset" fi echo "week ${pct:-?}%"
  2. Point Claude Code at it

    In ~/.claude/settings.json, set "statusLine": {"type": "command", "command": "~/.claude/statusline.sh"}. If you run several accounts with CLAUDE_CONFIG_DIR, each folder logs its own weeks.

  3. Read it at the end of the month

    Each file is one week. Average the last four for the seat's utilization. The number only updates while Claude Code is running, so a seat nobody opens has no files at all, which is itself the answer.

Measure agent wait time

Agents often stall on a person, not on the model. Claude Code's Notification hook fires when it sends a notification, with matcher values that include permission_prompt, idle_prompt and agent_needs_input.4 Log a timestamp from that hook and another when the next prompt is submitted; the gap is the wait. Sum it per week. A team whose agents wait hours a day for approvals has a process problem that no seat upgrade fixes.

A weekly review in 15 minutes

Weekly adoption review: one row per seat
ColumnWhat to writeAction if it is off
Active this weekYes or noTwo weeks of no: ask what is in the way
SessionsCountFew, short sessions: pair with a heavy user
Weekly limit used at resetPercentUnder 20% for a month: see the capacity guide; hits 100% early: consider a bigger seat
Wait timeHoursHigh: answer from phone, loosen safe permissions
Merged PRs with Claude CodeCountHigh use, no merges: look at review load and quality
A template by Orpheus, not a benchmark. Set targets from your own first month.

Pitfalls

  • Shared limits. Claude chat, Claude Code and Claude Desktop count toward the same limit,5 so a seat's weekly percentage includes chat. That is fine for a capacity metric; just don't call it pure coding.
  • Local logs miss other machines. Per-machine tools, including /usage, see only that machine.1
  • Rankings. The top-10 leaderboard is built for finding people who can teach others.3 Turned into a target, it gets gamed. See Should you rank engineers by AI token use?

Common questions

Can an admin see each person's weekly limit percentage?

Check your admin console. Anthropic's docs describe usage and contribution analytics, a spend report and, on Enterprise, an Analytics API. Each person can also read their own percentage in Settings > Usage or the statusline.

Does /usage show the whole team?

No. Its breakdown is computed from local session history on that machine and leaves out other devices and claude.ai.

Should we track tokens per engineer?

Track them for cost, not for performance. Token counts measure input, not results, and per-person rankings have been gamed. Pair any usage number with outcomes such as merged pull requests.

How long before adoption numbers mean something?

Give it about four weeks. Week one is setup, and a single week's weekly-limit percentage swings with deadlines. Compare four-week averages.

Sources

  1. Manage costs effectivelyClaude Code docs, undatedChecked
  2. Customize your status lineClaude Code docs, undatedChecked
  3. Track team usage with analyticsClaude Code docs, undatedChecked
  4. Hooks referenceClaude Code docs, undatedChecked
  5. How do usage and length limits work?Claude Help Center, undatedChecked