Copilot Usage Dashboard

A VS Code extension that shows token usage, AI Credits, and cost across GitHub Copilot Chat, Oh My Pi, and Pi sessions.

Install

npm install
npm run compile
npx @vscode/vsce package --allow-missing-repository
code --install-extension copilot-usage-dashboard-*.vsix

Open

Command Palette → Copilot Usage: Open Dashboard

Prompt-cache TTL

A cached prompt prefix is re-read at a fraction of the input rate, but providers expire it after a few minutes of inactivity — and once it lapses the whole conversation is re-billed in full. On a long agent session that is the difference between a few credits and a few hundred.

Enable copilotUsage.cacheTtl.enabled to get a live countdown of how long each session's cache stays warm, in the status bar, its tooltip, the sidebar, and the sessions table. Sessions still generating show HOT instead of a countdown.

Lifetimes are per-provider under copilotUsage.cacheTtl.ttl. Only Anthropic documents a TTL (~5 min), so treat the rest as tunable estimates rather than billing guarantees.

BYOK provider cost

Requests through your own endpoint — Azure AI Foundry, Anthropic, OpenAI, a local Ollama — are not billed by GitHub. They appear in the non-billable panel, where the credit columns show only what the traffic would have cost on Copilot: a figure on no invoice you actually receive.

The Provider cost column applies your provider's real per-token rates to the same traffic and reports USD. It is a different vendor's bill, so it is reported separately and never added to your AI credits.

Anthropic-family rates ship as defaults at published Claude API prices, which Anthropic documents as also being the Microsoft Foundry rate card. Override anything under copilotUsage.byokPricing, including regionMultiplier for Azure US Data Zone deployments (1.1x; Global Standard is 1.0x).

Two limits worth knowing:

• A model with no configured rate shows —, never $0.00. An unpriced Azure deployment and a genuinely free local model must not look identical.
• The total is a lower bound. Anthropic charges 0.1x for cache reads and 1.25x–2.0x for cache writes, but VS Code records one undifferentiated cached count. Cached tokens are priced as reads; copilotUsage.byokPricing.cacheWriteRatio shifts the split if your workload writes more than it reads.

Configure

Settings → search copilotUsage.

Common settings:

Full configuration reference, data sources, and internals: ARCHITECTURE.md

License

MIT. Portions of the prompt-cache TTL feature are derived from cache-timer (MIT © 2026 sukumarp2022) — see NOTICE.