Back to the collection
ContextPublic source

Token Weather: Claude Code context monitor.

Monitor Claude Code context usage above the prompt with Token Weather: a live percentage, token count, and 12-turn history.

By Claude Code DevRelReviewed 2026-10-07
Token Weather preview from the original creator

What it does

Token Weather is a source-published sample in Anthropic’s Claude Code playground. It draws a one-line context monitor above the prompt, using the session’s reported usage rather than a forecast of future requests. The display updates after each main-loop turn; it skips subagent turns.

  • Context percentage and input tokens against the model’s full window
  • Weather labels from Clear to Compact soon
  • Relative sparkline covering the last 12 turns
  • Change in tokens since the previous main-loop turn

How to install Token Weather

Use a terminal with Claude Code 2.1.287 or later, and read the author’s source and setup before loading it. The commands below clone the playground, check the plugin, then load it for one session. Validation checks the plugin; it does not certify code safety.

INSTALL / TERMINAL
git clone https://github.com/anthropics/claude-code-playground.git
cd claude-code-playground/claude-code/mods
claude plugin validate ./token-weather
claude --plugin-dir ./token-weather

For a persistent installation, run these author-documented commands from claude-code-playground/claude-code/mods, the same directory used above:

INSTALL / TERMINAL
claude plugin marketplace add ./
claude plugin install token-weather@claude-code-playground-mods --scope user

Start a new Claude Code session after installing. For loading errors, use the installation and display checklist. This listing follows the author’s terminal instructions; we have not tested the sample on Desktop.

What do the weather labels mean?

The sample labels full-window usage as Clear below 25%, Cloudy at 25–49%, Showers at 50–74%, Storm at 75–89%, and Compact soon from 90%. These are display thresholds, not predictions of when compaction will happen.

The token count is the input context for the last response, including uncached, cache-written, and cache-read tokens. It is not your remaining subscription allowance. The chart scales its bars relative to the fullest of the last 12 turns; use the percentage for absolute fullness.

Why does Token Weather disagree with the compaction notice?

They use different denominators: Token Weather uses the model’s full window, while Claude Code’s notice measures toward its lower auto-compaction point. The author’s screenshot shows 81% in the band alongside a 90% notice. This is a documented example, not a measurement by CC Mods.

Why does the reading stay at 0% or stop updating?

Before the first response reports usage, 0% is expected. The band refreshes after a main-loop turn, not during it or after a subagent turn. Starting a session or reloading the plugin resets its history. If the band is absent, check the terminal surface, plugin loading, and whether another mod draws in the same AbovePrompt band.

Does it save tokens or change compaction?

This sample displays usage; it does not reduce the context or change compaction policy. The author says its usage call requests no breakdown and sends no token-count request, and the mod itself calls no model. That describes this sample’s implementation, not a general savings claim for mods.

Source: Token Weather README, setup, and limitations. To compare other ways of monitoring a session, see the context collection and mods vs other extension types.

When this mod might be useful

Token Weather may be useful during a long coding conversation when you want context usage visible without leaving the prompt. Its turn history helps you notice changes in the session footprint. Use those readings as session information, not as a prediction of how many future requests you can make.

Understanding this listing

Use this listing as a starting point for comparing the author’s documented behavior with your own workflow. The repository is the source of truth for installation, configuration, and later changes. The commands here reflect the setup available at the review date.

Before you try it

  • The author documents Claude Code 2.1.287 or later and a terminal for this sample. No environment variables or configuration are required.
  • The percentage uses the full context window; Claude Code’s auto-compaction notice uses a lower threshold, so the readings can differ.
  • Updates once per turn, not continuously while Claude is working. The first reading can be 0% before a response reports usage.
  • History resets on session start or plugin reload. Another AbovePrompt mod competes for the same band.
  • Source and setup reviewed; CC Mods has not run the sample or verified compatibility on Desktop.