See how much of your Command Code plan you have left, without leaving the conversation.

Command Code plans (Go / GOAT / Pro / Max / Teams) pace your monthly credits with two rolling windows — a 5-hour cap and a weekly cap. A window opens on your first request and resets a fixed time later; it does not follow calendar days, and usage never carries over between windows. So "can I still finish this task?" cannot be answered from the monthly balance alone. What matters is how much of the current window is left and when it resets.

This plugin reads all three numbers and renders them where you are already looking.

Command Code · GOAT                             09-21 00:45 · 30d left in period
──────────────────────────────────────────────────────────────────────────
account  your-name <[email protected]>

5-hour window  ██░░░░░░░░░░░░░░░░░░░░░░   6.3%    $0.89 / $14.00
               resets 04:29 · in 3h 43m
weekly window  █░░░░░░░░░░░░░░░░░░░░░░░   2.5%    $0.89 / $35.00
               resets 09-27 23:29 · in 6d 22h
monthly        ░░░░░░░░░░░░░░░░░░░░░░░░   1.3%    $0.89 / $70.00
               $69.11 remaining

this period  330 requests · 100% success · 96.3M in / 306.9K out tokens
estimate      at your $0.0026 average, room for ≈ 5,064 more requests in the 5-hour window
              (based on your actual model mix this period; pricier models go much shorter)

When you are burning fast enough that a window will run out before it resets, it says so:

⚠ at the current $4.30/h, the 5-hour window runs out before it resets — exhausted in ~15m 20s

No hooks, no MCP servers, no agents, no background processes.

• A Command Code plan. Without one there is nothing to show. On pay-as-you-go rather than a subscription it still works, but shows a balance instead of windows.
• Node.js on PATH. The bundled scripts use only Node built-ins (node:fs, node:http, node:os, node:path, node:child_process, node:zlib) and the global fetch, so there is nothing to install.

Asking in plain language works too: "How much Command Code quota do I have left? Is it enough to finish what we are doing?"

Network: the panel calls four read-only endpoints on https://api.commandcode.ai — /alpha/whoami, /alpha/billing/credits, /alpha/billing/subscriptions, and /alpha/usage/summary. No other host is contacted. Nothing is sent anywhere else, and there is no telemetry.

The API key is resolved at runtime, in this order (first hit wins):

1. the environment variable COMMAND_CODE_API_KEY, CMD_API_KEY or COMMANDCODE_API_KEY;
2. ~/.commandcode/auth.json, written by logging into the Command Code CLI;
3. ~/.zcode/v2/provider_config.json — a provider whose api.baseUrl points at commandcode.ai, i.e. the key you already configured in ZCode.

The key is only ever placed in an Authorization: Bearer header. The plugin does not write, print, log or transmit the key, and it contains no credentials. --verbose reports which source was used (never the key itself).

• Reads the credential files listed above, and the API endpoints listed above.
• Writes nothing by default. No cache, no state, no config changes.
• --html writes one HTML file, only when you explicitly pass that flag, to the path you choose (default ./command-code-usage.html).
• --serve starts a local HTTP server on 127.0.0.1 (default port 8787) so a browser can poll the panel. It is off unless you pass the flag, binds to loopback only, and stops with Ctrl+C.
• The panel shells out to node <plugin>/scripts/cc-usage.mjs. That is the only process it starts.

"Room for ≈ N more requests" = remaining allowance ÷ your average cost per request this period. The average comes from your own usage, so the estimate adapts to any plan and any model mix without hard-coding per-model rates. Because the baseline is your own average, it stops holding the moment you switch models — the panel says so. Command Code's /provider/v1/models returns a model list with no allowance factors or prices, so "how many requests of model X specifically" cannot be computed from the API.

The warning extrapolates your current burn rate. That path has a trap: an hour after a window opens, extrapolating one hour of activity across seven days will always claim the weekly cap is about to blow — pure noise. So the script enforces a minimum sample: under 5% of the window elapsed it draws no conclusion at all. No warning therefore means "not enough data yet", not "you are safe".

Besides the command the agent runs (scripts/cc-usage.mjs), the plugin ships two standalone tools. Neither runs on its own; they exist for the cases described here.

• scripts/cc-usage.mjs — the panel itself. Runnable directly: node scripts/cc-usage.mjs --compact. The panel adds no state and no cache.
• scripts/install-user-scope.mjs — installs the commands and skill into your user-scope agent directories (~/.zcode/commands, ~/.zcode/skills) for users who prefer not to go through the marketplace. It writes those files and keeps a small manifest of what it wrote so it can update or remove them later. Do not run it on top of a marketplace installation: user-scope copies are discovered first and would shadow the installed plugin. --uninstall removes them.
• scripts/verify-discoverable.cjs — a read-only diagnostic that re-implements ZCode's own command parser, so a "my command does not show up" report can come with evidence. It reads files and prints a report; it writes nothing.

A custom command is ultimately a prompt: the body is injected, the agent runs the script, and the panel text passes through the model. Measured, one /quota costs roughly 390 tokens — about 100 for the command body, 80 for the tool call, 180 for the panel text, 40 for the reply. The body is deliberately short and the agent is told not to restate the panel, because the tool result is already visible. (The first version restated it and cost about 680.)

To spend no tokens at all, run the panel as a local page and open it in the agent's built-in browser pane:

node <plugin>/scripts/cc-usage.mjs --serve    # then open http://127.0.0.1:8787/

It refreshes every 30 seconds, shows the same ring gauges, and never touches the model.

ZCode exposes no plugin-contributed in-app widget, and a hook cannot display content either — the hook record it renders carries status, duration and name, with no output field. Inline shell expansion in command bodies (!cmd`` or a fenced ! block) does exist, but on Windows that shell is `cmd.exe` rather than bash, so it would need a hard-coded script path and would fail for marketplace installs. Not used here.

Issues and pull requests: https://github.com/Jovan1666/zcode-command-code-usage.

Not affiliated with Command Code. The plugin reads your own account's usage through the same endpoints the official CLI uses.

MIT