Marmot reads Claude Code session records stored on your machine and finds long contexts, idle MCP servers, cache misses, tool failures, and other causes of heavy usage.

Does Marmot upload my Claude Code sessions?

No. Marmot has no account or hosted service. Reports and session browsing stay on your machine.

How do I install Marmot?

Install the macOS menu bar app with Homebrew: brew install --cask drdroidlab/tap/marmot.

Claude Code · local diagnostics

Find the leak before Claude hits the limit.

Marmot turns local session records into evidence: what consumed your context, why it kept repeating, and what to change next.

Install for macOSHomebrew · macOS 14+

$brew install --cask drdroidlab/tap/marmot

See what Marmot finds ↘

Marmot

Usage

Max 20×

7 days30 days90 days

Daily modelled cost by model.

opus-5sonnet-5

10 Sep

11 Sep

12 Sep

13 Sep

14 Sep

15 Sep

16 Sep

Total · 7 days$507.48

Average / day$72.50

Peak day$308.46

ModelCostShare

opus-5$437.7186%

sonnet-5$69.6614%

Modelled at API rates — not your bill.

Field methodEvidence in.
Action out.

01Observe

Read the Claude Code session records already on your Mac.

02Trace

Connect heavy usage to context, tools, MCP, and model choice.

03Act

Get one concrete change to make before the next turn.

Follow the token trail

Most waste looks harmless until you add it up.

Marmot turns vague usage pressure into a short list of causes, with evidence attached and an action you can take.

Context

The task ended. The context kept travelling.

Old turns and tool results ride with every new request. Marmot shows how much history each turn is carrying.

Evidence450K history · 83 prompts

450Ktokens resent / turn

Next moveCompact the session, or start the next task fresh.

MCP

Loaded on every request. Called zero times.

MCP definitions consume baseline context. Marmot separates useful servers from idle ones.

Evidence17,597 tokens from servers never called

workspace108× called

database4× called

issue-tracker0× called

Next moveDisable the idle servers before the next session.

Tools

A failed call often buys you another expensive turn.

See which tools fail and where retries are quietly multiplying usage.

Evidenceexecute_sql · 3 of 3 calls failed

Tool calls11,249

Bash3% failed

execute_sql100% failed

Next moveFix the tool contract instead of paying for another retry.

Models

Not every job needs the most expensive model.

See where premium models handled work a lighter model could carry.

EvidenceOpus · 86% of modelled spend

Opus86%

Sonnet14%

Next moveRoute routine work to a lighter model.

Not another usage meter

Your limit says what is left.
Marmot says what to fix.

Recommendations point to the sessions, context, and habits behind your usage—then give you a concrete next move.

Marmot

UsageRecommendations

Recommendations

From the last 7 days of Claude sessions on this Mac. Thresholds you set come first, then the biggest savings.

Long session · 3 sessions

23 prompts in this session so far, never compacted.

↳ Start a fresh session for a new task, or run /compact to keep the thread and drop the weight.A threshold you set in Notifications

Each turn re-sends 450K tokens of history

Across 83 prompts in one long-running session.

↳ Run /compact, or start fresh for the next distinct piece of work.Measured from your local session files

95% of usage was at >150K context↳ Compact when the early part stops being relevant.Based on Claude Code usage data

78% of usage came from 8+ hour sessions↳ Start a fresh session at the next piece of work.Based on Claude Code usage data

Inside the local browser

Go from a number to the thing behind it.

Open a detailed report on your Mac when you want to dig deeper.

01 / OVERVIEW

Know which limit is actually close.

See plan windows, modelled spend, and session activity together—without mistaking the shadow price for a bill.

Marmot · local report

What is running out

5-hour session3%resets in 1.7h

Weekly limit2%resets in 6.4d

Modelled spend$3,255at API rates

Sessions114last 30 days

02 / TOOL FAILURES

Spot the retries quietly adding up.

Find tools that fail repeatedly so you can fix the path, permission, or contract before the next run.

Marmot · local report

Tools failing

Failures cost a turn—and often a retry.

Bash

292 · 3%

browser_batch

6 · 13%

execute_sql

3 · 100%

03 / MCP CONTEXT

Find the servers you never called.

See which attached servers were useful and which definitions travelled with every request for nothing.

Marmot · local report

MCP servers

17,597 tokens from idle servers.

workspace

108× called

database

4× called

issue-tracker

never called

monitoring

never called

Nothing leaves your machine.

Local by design

Marmot goes through your sessions. Your sessions go nowhere.

No account. No hosted service. No telemetry. The normal report avoids conversation text, and the local browser can leave prompts and replies out too.

Reads~/.claude/projects/**/*.jsonl
Runs onmacOS 14+ · Node included
AuditableOpen source · MIT licensed

Install Marmot

One command. Marmot in your menu bar.

Install the macOS app with Homebrew and keep your limits, cost, and recommendations one click from the clock.

FAQ

Before Marmot starts digging.

What exactly does Marmot read?

The JSONL session records Claude Code already keeps under ~/.claude/projects. The standard report uses counts, IDs, models, tool names, and usage data—not your conversation text.

Is modelled spend my actual bill?

Not on a subscription. It is what the tokens would cost at published API rates. The session and weekly limit percentages are the useful ceilings on Pro, Max, and Team.

Will Marmot interrupt every turn?

No. A rule speaks once per session. Limit and long-session reminders return only at their configured marks, with a quiet period between live nudges.

Can I browse the exact session behind a number?

Yes. Run marmot browse to open a local session browser. Use marmot browse --no-text to exclude prompts and replies.

Is Marmot stable?

Marmot is alpha. Claude Code’s session format is internal and can change. Run marmot doctor when something looks wrong.

There is probably a reason your limit disappeared.

Let Marmot find it.

Install Marmot Star on GitHub ↗

Find the leak before Claude hits the limit.

Usage

Recommendations

Most waste looks harmless until you add it up.

The task ended. The context kept travelling.

Loaded on every request. Called zero times.

A failed call often buys you another expensive turn.

Not every job needs the most expensive model.

Your limit says what is left.Marmot says what to fix.

Recommendations

Go from a number to the thing behind it.

Know which limit is actually close.

What is running out

Spot the retries quietly adding up.

Tools failing

Find the servers you never called.

MCP servers

Marmot goes through your sessions. Your sessions go nowhere.

One command. Marmot in your menu bar.

Before Marmot starts digging.

Let Marmot find it.

Your limit says what is left.
Marmot says what to fix.