ClaudeMeter
New · Claude Meter Pro · Waitlist open

A background agent that quietly cuts your Claude usage.

Claude Meter Pro is a proactive AI agent that runs alongside Claude Code and Claude.ai, watches what your sessions actually spend, finds the leaks, and patches them for you (with your permission) so you keep shipping after everyone else hits the wall.

Never hit “Model token limit reached” again.

No spam. We email when your invite is ready and that's it.

Local-first, your code never leaves your machineQuality-gated, every fix runs a shadow evalReversible, every change is logged
How it works

Watch. Suggest. Verify. Ship.

Pro is a four-step loop that runs in the background. You stay in flow, the agent does the cleanup.

  1. 01Watcher

    Reads what your Claude sessions actually spend

    The agent watches your Claude Code calls and Claude.ai chats locally: which tools you call, how big the context windows get, where you re-read the same files, where Opus runs when Sonnet would have been enough. No code leaves your machine.

  2. 02Suggests

    Real-time, in-context savings suggestions

    When the agent spots a fix, e.g. trim a 40k-token system prompt to 6k, swap a model class, cache a recurring file read, batch tool calls, it surfaces the suggestion right in the menu bar with the projected percentage saved on your 5-hour and weekly buckets.

  3. 03Reports

    Weekly report on what you almost wasted

    Every Monday morning you get a one-screen breakdown: where the week's tokens actually went, the top 3 leaks, the savings the agent already booked, and the ones still waiting on your approval.

  4. 04Auto-fix

    Opt-in: it ships the fixes for you

    Flip a toggle and the agent stops asking permission for low-risk fixes (prompt trims, file de-duplication, model swaps inside the same family). High-risk fixes still wait for one-tap approval. Every change is reversible, every change is logged.

  5. 05Quality-safe

    Quality gate, not just a token counter

    Each suggested change ships with a tiny shadow eval: the agent re-runs your last few real prompts against the proposed setup and only ships the fix if output quality stays inside your tolerance band. Saving tokens that break your work isn't saving anything.

  6. 06Local-first

    Local-first, MIT core, no telemetry

    The agent runs on your machine alongside the existing Claude Meter menu bar app. Pro is a paid layer on top of the same MIT codebase. Your prompts, code, and reports stay local; only billing pings the Pro server.

What it catches

The leaks burning your weekly cap on a Tuesday.

Bloated system prompts
40k-token preambles trimmed to the 5k that actually changes the answer.
Re-read context
Same file dumped into 12 sub-agents in one session, deduplicated to one read.
Opus where Sonnet wins
Per-task model class downgrade when the quality eval says nothing changes.
Tool-call fan-out
10 sequential greps coalesced into one parallel call, two messages instead of ten.
Loop transcripts
Long agent loops summarized into a 2k-token recap before the next step.
Prompt-cache misses
Stable prefixes pinned so the cache hits instead of re-billing every turn.
FAQ

Common questions before you join the waitlist.

How is this different from the free Claude Meter?
Free Claude Meter is read-only: it shows you the server-truth 5-hour and weekly quota in your menu bar. Pro is the active layer on top: it watches what you actually spent on, suggests cuts, and (with your permission) implements them so the bar stops climbing as fast.
Will it touch my code without asking?
Not unless you flip the auto-fix toggle, and even then only for the categories of fix you whitelist. By default, every change is a one-tap approval. Every applied change is logged and reversible.
How does the quality gate work?
Before applying a fix, the agent re-runs your last N real prompts against both the current setup and the proposed setup, scores them with a held-out judge, and only ships the change if the quality delta is inside the tolerance you set. Saving 30% of tokens is useless if the answers got worse.
When does it ship?
We're inviting waitlist signups in batches over the next several weeks. Earliest invites go to people who already have the free menu bar app installed. Drop your email and we'll email when your slot opens.
What does it cost?
Pricing is finalized at invite time and locked in for waitlist signups. We're targeting a flat monthly rate that pays for itself the first week the agent prevents one weekly-cap wall.

Get on the waitlist.

Pro invites go out in batches starting with people who already run the free menu bar app. Drop your email and we'll send your slot when it opens.

No spam. We email when your invite is ready and that's it.

Want the free menu bar app first? Install Claude Meter or back to the homepage.