How one tool tallies AI coding usage per person across Cursor, Claude Code and Codex

The question no single dashboard answers

Few coding teams stick to one AI assistant. One person works in Cursor on the desktop, another runs Claude Code or Codex in a terminal, and a third keeps GitHub Copilot open in the editor. Devin, another coding agent, can run either as a desktop app or as a session in the cloud.

As dev.to reported, each vendor shows usage for its own product, but a manager wants something different: which employee used which agent, on which surface, and how many tokens that prompt consumed. Tokens are the small chunks of text that AI models read and write, and they are the usual unit for what an AI request costs.

The post describes a tool called AI Usage Metrics as one way to fill that gap. It is a write-up about a single product, so treat it as one approach, not a verdict on how every company should measure.

What goes into each row

The idea is simple. When an employee finishes a prompt, the system records one row. That row holds the person, the agent, the surface, and three token counts: input (what was sent to the model), output (what came back), and cached (input the model had already seen and could reuse more cheaply).

The prompt itself never leaves the employee's computer. Only the counts are reported. A prompt that was skipped produces no row, and a later prompt never overwrites an earlier one. The number of turns is just the number of rows.

Cursor, Claude Code, Codex and Copilot can each send these rows automatically from the editor through a hook, which is a small trigger that fires after a prompt. Devin Desktop can send the same kind of row.

Why the surface matters

The tool sorts every row into one of three surfaces. Desktop covers Cursor, Copilot in the editor and Devin Desktop. CLI, short for command-line interface, covers Claude Code and Codex running in a terminal. Cloud covers a Devin session that ran remotely.

This split matters most for Devin. It is one agent used in two very different ways. If a company only counts "Devin," it blends someone working at their own machine with a job running in the cloud. A row that names the surface keeps them apart.

With those rows, a company can total tokens per employee, per agent and per surface, count completed prompts, and work out how much input came from the cache. A high cache share suggests a lot of repeated context is being reused.

Two units, one alert

Not everything is counted in tokens. Devin's cloud usage arrives once a day from the company's enterprise account and is measured in ACUs. The post says that figure is not added to the token columns used for the other tools. Copilot's daily report is measured in AI credits instead.

On top of the per-person totals sits a combined usage alert. It adds up one person's activity across Cursor, Claude Code, Codex, Copilot and Devin. The default cap is 250 USD over 7 days, and companies can change both numbers, along with the currency and time zone.

Someone at or over the cap gets a single email per local day. People with no email address on file are skipped. Emails go out during the hours when that company's usage is lowest, and a jobs page records each run, including how many people were mailed and whether it finished.

Setup, health checks and limits

Signing in requires a company email, and personal inboxes such as Gmail are declined. The company's identity is its email domain. In settings, an admin pastes a Devin enterprise key and, for Copilot, the GitHub enterprise name and a token. Saving either turns on a daily pull. Secrets are entered once and not shown again. A third credential for repository access can be stored but does not start a pull.

The daily pull runs once per company per UTC day. Devin pulls sessions updated the previous UTC day, while Copilot pulls a report from two UTC days earlier. A missing key for one does not block the other.

A jobs page also reports the health of each hook connection with four states: nothing received yet, a message within the last day, silence for more than a day, or a failed last call. The daily Devin and Copilot pulls are separate from their desktop and editor hooks, so one does not mark the other as working.

The post is explicit about limits. Vendor documentation remains the authority on any one product, and this report does not replace the vendors' own consoles. It lines people up across them. The public demo uses sample data.