Tokendraw — AI skill for Claude Code

tokendraw

Tokendraw is a command line tool for estimating the energy usage and emissions of LLM coding agent use.

Tokendraw currently supports Claude Code, Codex, pi, as well as custom adapters for other tools. It reviews the logs of these tools, counts the tokens of model calls, estimates the electricity used to serve those calls, and the emissions from producing that electricity. The embodied emissions of hardware, training or water use are not estimated yet.

Getting started

npm install -g tokendraw

or `cargo install --locked tokendraw` (see [Building from source](#building-from-source)). Then:

tokendraw report

The first run reads all your agents' logs, which can take a minute. The coverage table shows where tokendraw looked for each agent's logs - if it says `not found` for an agent you use, see [Where the logs are](#where-the-logs-are) for the variable that points tokendraw to them.

On Windows, use PowerShell: the commands tokendraw suggests are quoted for it.

Next, run `tokendraw projects` to assign calls that belong to no project (see [Projects](#projects)), and if the report names agents or models tokendraw cannot read or has no data for, [send us your setup](#sending-us-your-setup).

Configuring for your setup

Where the logs are

Tokendraw reads each agent's logs where the agent writes them, from the same environment variables:

• Claude Code: $CLAUDE_CONFIG_DIR/projects, else ~/.claude/projects
• Codex: $CODEX_HOME/sessions, else ~/.codex/sessions
• pi: $PI_CODING_AGENT_SESSION_DIR, else $PI_CODING_AGENT_DIR/sessions, else ~/.pi/agent/sessions

Sessions pi keeps elsewhere through its `--session-dir` option or `sessionDir` setting are not read.

Projects

A call belongs to the project of the Git checkout its working directory is in, including a directory that was removed if tokendraw saw it while it existed. Calls it cannot place, such as those of worktrees removed before tokendraw first ran, stay