AI Cost Coach for VS Code

See where your AI spend goes, from a month down to a single tool call, for Claude Code, GitHub Copilot CLI and VS Code Copilot Chat. A coach holds the prompts that would waste money, and a daily budget keeps your Copilot premium requests in check.

The screens are in the tour: run AI Cost: Show Tour from the Command Palette, or open Help → Get Started → AI Cost Coach. They are made-up sample sessions, so nothing in them comes from anyone's real logs.

The screens

Run AI Cost: Show Tour (or Help → Get Started → AI Cost Coach) for a walkthrough. The extension's page in the Extensions view shows this text without the images, because VS Code only loads images there from public https addresses.

• Overview. Opens in an editor tab: tiles, cost by month stacked by tool, tokens, and cost by source, model and prompt. By month · By day · 7d · 30d · 90d changes the range. The legend is the tool selector: click a tool to leave it out (the same filter as Filter → Source, so the tree, Advice and status bar follow). Almost everything is clickable: a bar filters or opens that month, day, model or prompt.

• The sidebar has four panels: a compact Overview, Sessions (a tree: session → prompt → model call), Advice and Coach.

• A month. Click a month to see its days; a breadcrumb takes you back.

• A session shows its cost per prompt, where its tokens went, its models and every prompt.

• A prompt shows its model calls, cache hit, token split and tool calls, with a breadcrumb and Back button.

• A model call shows what went in (what was new since the previous call) and what came out (reply, reasoning where recorded, tool requests).

• A tool call shows its full input and result, with status, size and time.

Filter (the funnel in the Sessions title bar) narrows by source, day, model, minimum prompt cost and text; Sort orders by time, cost, tokens, cache hit or calls; Sync reads the logs again (they are also re-read every minute).

Copilot premium requests and the daily budget

Under the tiles, a bar shows your plan's premium requests used and left this month, the reset date, and where the month will end at this pace. The numbers are the ones Copilot CLI writes into its own log, so nothing needs configuring (Claude Code has no such quota, and the panel is hidden on unlimited plans). The status bar carries it: AI today $1.20 · Copilot 6,832 left (89.9%) · 3/16/day, and turns to a warning colour when the month is on pace to go over or today is over budget.

The daily budget is what is left spread over the weekdays remaining (none at weekends), or your own number: Daily budget (premium requests) in the Coach panel, or cost coach budget 50 in a terminal (budget auto goes back). Today's use is what this machine's Copilot sessions used since midnight. In strict mode the first prompt after it is reached is held, once a day.

The model guard says so once when requests are on pace to run out, naming the models that cost you the most requests a prompt and the cheaper one you already use most. It advises only: it never hides or switches a model.

Advice

The Advice panel measures your own sessions and lists where spend could be cut, each with an estimated saving and the prompts behind it: prompt cache expired between prompts, model switched mid-session, top-tier model on short replies, very long conversations (over 150k tokens of context), and a few prompts carry the spend. Savings overlap, so they do not add up. Under them, unused habits are listed with no saving of their own: /compact never used in a long session, no project instructions file, a token logger that is not recording, or cost tips switched off.

Copilot tracking and cost tips

Without setup, Copilot sessions already show credits, dollars and prompt tokens from Copilot CLI's log and VS Code's chat storage. Copilot tracking adds per-call tokens (a small Copilot CLI extension records every model call; prompts it did not see show "No token data") and the coach, a prompt hook that checks each prompt before it runs in Copilot CLI and VS Code chat.

Turn it on with Copilot tracking in the Coach panel, or AI Cost: Set Up Copilot Tracking. It asks first and lists what it writes: the hook (~/.copilot/hooks/ai-cost-drilldown.json), the token logger (~/.copilot/extensions/ai-cost/) and the hook's scripts (~/.ai-cost-drilldown/runtime/). Copilot CLI loads extensions only in experimental mode: run /experimental on once, then start a new session. Remove Copilot Tracking deletes the three.

The switches (Coach panel; set by environment variables such as AI_COST_COACH_STRICT when those are present): Coach (a one-line tip when a prompt is about to cost more than it needs to, on), Strict mode (hold the prompt with a banner instead; send the same prompt again to run it, on), Hold on every cache expiry (on), Strict model (the banner also carries the prompt reworded by a small model, off) and Model guard (on).

A banner says why, gives a Tip on what to change, and lists every rule that applied (Also:). With strict mode off each is a one-line tip. Questions, "yes", follow-ups, anything naming a file, ticket or number, "fix the failing tests", /compact, /clear, /new and the prompts Copilot sends by itself are never held.

In VS Code chat a blocked prompt would just vanish, so the hook blocks only while this extension is running to show why. Each hold is a notification: Why (the full banner), Send anyway, Copy prompt. With no window open, or aiCost.notifyHeldPrompts off, you get a tip instead.

Coach results shows the last 30 days: tips shown, prompts held, whether you changed the prompt or sent it anyway, and an estimated cost avoided (list prices, not a measurement). The limits (Cost limit, Long session, Daily budget, Big paste, Model guard: costly from) are editable there, and Pause strict mode stops holds for a while.

The coach learns your cache lifetime from the gaps in your token log (it overrules the documented 5 or 30 minutes only with five or more telling gaps) and checks its cold-cache warnings against the next call. Both need the token logger.

If the hook does not run. VS Code starts hooks with a thin PATH, so a bare node fails and prompts run unchecked. Setup writes the absolute path to node, and AI Cost: Check Copilot Tracking runs the hook the way VS Code does and tells you whether it works (it also runs after setup and at startup). A failure inside the hook is written to ~/.ai-cost-drilldown/hook-errors.jsonl. On Windows the hook file names a small hook.cmd that runs from cmd.exe, PowerShell and Git Bash alike.

Reading the numbers

Costs are estimates. Claude Code's cost is each model's list price applied to its transcript's tokens; Copilot's is the AI credits it recorded, at $0.01 per credit, with each call priced at GitHub's published token prices. Subscription usage is not billed per token: these are list-price figures for comparing where usage went. Cache: every call re-sends the whole conversation, normally from a cheap cache; a prompt that starts with a miss pays full price for all of it. Sources: ~/.claude/projects/, ~/.copilot/session-state/ and VS Code's chat storage.

Settings and commands

Where to change things: the settings below are in VS Code Settings (search AI Cost, or aiCost. in settings.json). The coach's switches and limits (strict mode, the daily budget and the rest) are in the Coach panel in the sidebar, and are saved in ~/.ai-cost-drilldown/config.json. Commands are in the Command Palette under AI Cost: ….

From a terminal, the same switches are node ~/.ai-cost-drilldown/runtime/bin/cost.js coach [on|off|strict on|off|budget <n>|auto] (it writes the same config.json, so the Coach panel shows the change). Only the extension's own commands are on the Command Palette: there is no cost command on your PATH.

Your data and privacy

The extension reads session logs that already exist on your machine, which include your prompt text, folders and branch names. It sends nothing anywhere and needs no account or key. The one exception is Strict model, which sends the held prompt, the session's last few prompts and answer, its last failed tool result and the branch name to a small model through your own copilot command. It writes nothing to those logs. Its own files are in ~/.ai-cost-drilldown/ (AI_COST_HOME changes it); Copilot tracking is the only thing that writes outside it, and only after you agree.

What it cannot do

• Recover tokens it did not record, or give per-call detail for VS Code chat (one total per request).
• Show Claude's reasoning or Copilot's full replies.
• Go back past what the tools keep (Claude Code deletes transcripts after about 30 days unless cleanupPeriodDays is raised).
• Give bill-accurate totals: it reports what the tools logged on this machine.
• Coach Claude Code (its sessions are measured; tips and strict mode act on Copilot only), or hold a prompt the hook never saw. If the hook cannot run, Copilot lets the prompt through, which is why Check Copilot Tracking exists.

Checked live: strict mode in a Copilot session inside VS Code. Not yet checked: the classic Copilot Chat panel, a hold in a live Copilot session on Windows, and the token logger across Copilot CLI updates.