Adapterly Monitor

Track your Claude Code spend. Locally.Beta

A tiny CLI that reads your ~/.claude/projects/ session logs, aggregates spend, and surfaces savings ideas. Nothing leaves your machine.

Install

curl -sSf https://adapterly.ai/monitor/install.sh | bash

The installer:

Checks Python 3.10+ is available
Installs pipx if missing
Installs adapterly-monitor from a hosted wheel (no PyPI publish yet during beta)
Registers the MCP server in Claude Code (if the claude CLI is present)
Runs a first sync and prints your weekly report

Prefer manual? pipx install https://adapterly.ai/monitor/adapterly_monitor-0.1.0-py3-none-any.whl, then adapterly install-mcp if you use Claude Code.

Use from the terminal

adapterly today Today's spend + top models adapterly week Last 7 days, per-day breakdown adapterly month Last 30 days adapterly projects Spend per project (Claude's cwd) adapterly models Spend per model, cache stats adapterly sessions Most expensive sessions in the window adapterly recommend Savings suggestions from your usage adapterly watch Background daemon + desktop notifications adapterly sync Force re-read of the JSONL logs

Use from Claude Code chat

After install, restart Claude Code. The MCP tools appear in-chat — ask directly:

you →

"how much did I spend on Claude today?"

you →

"which project is my biggest cost this week?"

you →

"give me three concrete ways to reduce my Claude spend, based on my actual usage."

Available MCP tools: cost_today, cost_window, cost_by_project, recommendations.

What it looks like

$ adapterly recommend
Recommendations  (based on last 30 days)

1. Opus is 100% of your cost — try Sonnet-first routing  ~$166/mo  [medium effort]
   Most Claude Code prompts don't need Opus-tier reasoning. A common pattern is:
   default to Sonnet, escalate to Opus only when the response is flagged
   low-confidence or the task explicitly requires deep reasoning.

2. Your most expensive session cost $518.55 — long context accumulates fast  ~$156/mo  [small effort]
   Session `7c214da7` in /home/you/repo accumulated 715 calls. Very long
   sessions blow the cache window and force expensive re-reads. Consider
   `/clear` mid-way or splitting the work across sessions.

3. Cache-read ratio only 24% — you're paying full price for repeated context  ~$40/mo  [small effort]
   Prompt caching costs ~10% of full input pricing after the first hit.
   Claude Code caches system prompts automatically; explicit `cache_control`
   markers in your skills and MCP tools lift the hit rate.

Privacy

Everything runs locally. Data stays in ~/.adapterly-monitor/db.sqlite.
No telemetry, no cloud sync, no analytics call-home.
The MCP tool runs inside your own Claude session — the data you see, Claude sees. No other party.
Uninstall: pipx uninstall adapterly-monitor && rm -rf ~/.adapterly-monitor.

Status

Beta. Model pricing is illustrative — verify against Anthropic pricing before quoting it to your CFO. Feedback + bug reports: hello@adapterly.ai.

This is a companion to Adapterly's consulting offering. If you want the same analysis but for a whole team or a production workflow, book an AI Cost Review.