Coding agents

Claude Code

Add this to ~/.claude/settings.json, or to Claude Code's managed settings to roll it out to everyone. Delta is Claude Code's default.

{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_METRICS_EXPORTER": "otlp",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/protobuf",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "https://<your ingest address>/v1/otlp",
    "OTEL_EXPORTER_OTLP_HEADERS": "Authorization=Bearer <your connector token>",
    "OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE": "delta"
  }
}

Gemini CLI, Copilot Chat and other tools, through a Collector

Run the contrib OpenTelemetry Collector where the tools run and point each tool's OTLP export at http://localhost:4317 (gRPC) or :4318 (HTTP). The Collector adds the token, sends HTTP to Token Meter, and converts cumulative metrics to delta.

receivers:
  otlp:
    protocols:
      grpc: { endpoint: 0.0.0.0:4317 }
      http: { endpoint: 0.0.0.0:4318 }
processors:
  cumulativetodelta: {}
  batch: { send_batch_max_size: 5000 }
exporters:
  otlphttp:
    endpoint: https://<your ingest address>/v1/otlp
    headers:
      Authorization: "Bearer <your connector token>"
service:
  pipelines:
    metrics:
      receivers: [otlp]
      processors: [cumulativetodelta, batch]
      exporters: [otlphttp]

Kiro (Enterprise)

Kiro sends from AWS once a day at 02:00 UTC, for the previous day. Store this as the plaintext of an AWS Secrets Manager secret encrypted with a customer managed KMS key, give q.amazonaws.com read access to the secret and the key, then pick the secret in the Kiro console under OpenTelemetry export with the HTTP/protobuf protocol.

{
  "OTEL_EXPORTER_OTLP_ENDPOINT": "https://<your ingest address>/v1/otlp",
  "OTEL_EXPORTER_OTLP_HEADERS": "Authorization=Bearer <your connector token>"
}

A resent day replaces the stored day, so Kiro's retries never count twice.

Cline (Enterprise)

In the Cline dashboard's Remote Configuration, turn on OpenTelemetry with logs enabled and set these. Cline reports usage as log events; metrics can stay off.

OTLP endpoint:  https://<your ingest address>/v1/otlp
Protocol:       HTTP/protobuf
Header:         Authorization: Bearer <your connector token>

Continue

Add this to Continue's config.yaml, or to a shared assistant on Continue Hub. Continue posts one tokensGenerated event per model call, and no code is sent at level noCode.

data:
  - name: Token Meter
    destination: https://<your ingest address>/v1/connectors/continue
    schema: "0.2.0"
    level: noCode
    events: [tokensGenerated]
    apiKey: <your connector token>

Say who is using a tool

Usage is linked to a person by user.email, then user.account_uuid, then the tool's anonymous user.id. A tool that reports no user can be attributed by setting a resource attribute where it runs. Without it, usage still counts but is not linked to anyone.

export OTEL_RESOURCE_ATTRIBUTES="user.email=you@company.com"

What the endpoint accepts

• OTLP over HTTP only, as application/x-protobuf or application/json, gzipped or not. Tools that only speak gRPC go through a Collector.
• Delta temporality only. Cumulative points are rejected with a message asking for OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE=delta; a Collector with cumulativetodelta converts them.
• At most 5,000 data points per request.
• Points dated more than 400 days ago or more than a day ahead are rejected.
• Metrics the endpoint does not map are ignored, not treated as errors.

Agent usage

Developers › Agent usage shows how teams use AI coding tools over 7, 30, 90 or 180 days or a year: active people, agent sessions, AI commits and pull requests, pay-as-you-go spend and API-equivalent value; agent sessions by tool, prompts sent by tool, active people by type of engagement (terminal, chat, agent, web); lines, commits and PRs per day; and By tool and By team tables. Select a row for every figure, including tokens and accept rate.

• Sessions and prompts are separate charts because they are different units. Claude Code, Gemini CLI and Cline report sessions; Cursor, Copilot, Codex and Claude chat report prompts.
• With GitHub connected, the daily output charts come from GitHub and split human from each AI tool. Human is an upper bound: a commit without an AI marker counts as human.
• With Cursor AI Code Tracking, a separate chart shows exact committed lines written by Cursor against everything else. Only that tool is measured, so it is not a workspace-wide AI share.
• Spend is pay-as-you-go where the tool bills it, otherwise API-equivalent value (subscription use valued at API rates, not money spent), otherwise credits. These are never added together.
• Teams under 5 people are combined in By team so no individual can be read from a row.
• A connected source that needs attention (a rejected key, a missing workspace ID or organization login) shows a notice above the charts with a link to fix it.

Team

Developers › Team (team workspaces) has an Overview tab with the org adoption rate, daily active AI users, power users and average cost per active user, plus an Attribution card and a Coding agents card showing spend per tool and how much of it is linked to people. The Teams and Roles tab has the team matrix, power users by team and usage by role. Admins can set a target share of active people per team with Edit targets. Active means at least one SDK or connected AI-tool activity in the range.

Individual

Developers › Individual has two views. People (team leads and above) lists AI spend and engineering output per person, with no score or ranking; select a person for their daily tokens by tool, coding tools and SDK-metered models. Me is every member's own usage, output and Claude plan, and is visible only to them.

Your agents

Developers › Your agents (admins, team workspaces) is the inventory of AI tools and accounts: Tools in use (sanctioned and shadow tools), Discovery (Google Workspace and expense or card CSV uploads, which feed Shadow AI signals), Registry (AI assets and SaaS agreements; register a tool by hand) and Seats and licences (seat utilization, idle seats, renewals, and vendor admin connectors for GitHub Copilot seats and Slack).

Seat management

Developers › Seat management (admins, team workspaces) measures how much of each Claude seat's weekly allowance is used. Anthropic publishes no token or dollar quota for a seat, so a seat's use is reported when Claude showed a percentage, a calibrated estimate when usage is divided by an allowance learned from readings, or a peer estimate otherwise.

1. Add plan: name, type (Max / Pro, Team or Enterprise) and, for Enterprise, seat-based or usage-based billing. Each person belongs to exactly one plan.
2. Add a seat: one email and tier, or import a CSV of email,tier. People who use Claude but hold no seat appear under Not on any plan, where you can assign them.
3. Set each tier's Seat price / month and, optionally, declare its Weekly allowance.
4. Add Usage readings: the weekly percentage a seat holder sees in Claude under Settings, Usage. Readings above 50% are the most accurate; one below 30% is ignored.
5. With a directory synced (Google Workspace, Okta or Entra), Add a directory rule selects people by organisational unit, department or job title. You review them before anyone is added.

Seats are graded over the last 4 complete weeks as Achiever, Steady, Laggard, Dormant or Unknown, with "at cap" and "downgrade?" flags. Values are API-equivalent dollars of Claude Code usage, not money spent; chat is not included unless a feed reports it. Usage comes from the Claude Code feeds above (Admin API, Enterprise Analytics, OpenTelemetry or AIScale), and with Claude Enterprise connected each plan card also shows how many seats Anthropic reports as assigned. Every view of the named seat list is recorded.

On Individual › Me, a member sees their own seat and weekly use and can Add a reading. Pro and Max subscribers can install a Claude Code status-line reporter, shown on that tab, that forwards only the weekly percentage and their email.