Context Lens

Cost intelligence and context optimization for AI coding assistants. Works with Claude Code, Cursor, and any tool that runs Claude models.

Track spending, audit context bloat, disable unused skills and MCPs, and run leaner sessions. Automatic logging, context analysis, optimization recommendations, budget alerts, and session profiles — all local, zero dependencies.

Why Context Lens

Every skill, MCP server, and plugin you install inflates your context window — and you pay for that context on every single message. A typical setup with 60+ skills and 10 MCP servers burns ~145K tokens of overhead before you even ask a question.

Context Lens shows you exactly what that overhead costs, which pieces you actually use, and lets you disable the rest per-session.

Features

• Automatic logging — SessionEnd hook silently records cost data every time you exit a session
• Cross-session tracking — Persistent JSONL log accumulates data across all terminals and projects
• Session intelligence — Captures /rename names, auto-classifies sessions (debug, build, review, refactor, test, docs, deploy, config)
• Project detection — Auto-identifies projects from git remote URLs
• Model comparison — Shows what your work would cost on Opus, Sonnet, and Haiku
• Subagent & cache tracking — Includes subagent tokens and cache hit ratios
• Real-time monitoring — Cost tips when context or spend thresholds are crossed
• Budget alerts — Daily/weekly/monthly limits with 80% and 100% warnings
• Weekly digests — Automatic week-over-week comparison on session start
• Context snapshots — Logs installed skills/MCPs at session start for trending
• CSV export — Dump all data for external analysis

Installation

/plugin marketplace add ChewbaccaRoars/claude-code-cost-tracker
/plugin install cost-tracker@claude-code-cost-tracker
/reload-plugins

Or from a local clone:

git clone https://github.com/ChewbaccaRoars/claude-code-cost-tracker.git
# Then in Claude Code:
/plugin install --path ./claude-code-cost-tracker

Skills

/cost-tracker — Spending Reports

Command	Description
/cost-tracker	Today's cost summary
/cost-tracker week	Last 7 days
/cost-tracker month	Last 30 days
/cost-tracker all	Full history
/cost-tracker compare	Model cost comparison only
/cost-tracker export	Export all data to CSV
/cost-tracker project:<name>	Filter by project name
/cost-tracker session:<name>	Filter by session name

Reports include token breakdowns, per-project costs, model comparisons, cache efficiency, and daily trends.

/cost-audit — Context Bloat Analysis

Scans your entire setup and shows what's inflating your context window:

Context Audit — Last 30 days (147 sessions)

TOTAL CONTEXT LOAD: ~145K tokens/session ($0.043/msg on Opus)

TOP WASTE — Skills not earning their context cost:
  #  Skill              Size     Last Used    Usage   $/Month  Action
  1  slide-icons        23KB     64 days ago  0.7%    $2.32    DISABLE
  2  workday            21KB     45 days ago  1.3%    $2.15    DISABLE
  ...

MCP SERVER WASTE:
  #  Server          Tools  Last Used     Usage   $/Month  Action
  1  atlan-mcp        18    31 days ago   0.0%    $1.08    DISABLE
  ...

ESTIMATED MONTHLY SAVINGS IF ALL "DISABLE" APPLIED: $14.20/month

Verdicts:

• KEEP — Used in >10% of sessions or used in last 7 days
• REDUCE — Used occasionally but oversized (>10KB)
• DISABLE — Not used in 30+ days, costing money for no value

/cost-lean — Session Context Profiles

Manage context profiles to run lean sessions:

Command	Description
/cost-lean	Show current context load with cost per item
/cost-lean recommend	Generate profiles from your usage patterns
/cost-lean save <name>	Save current enabled config as a profile
/cost-lean profiles	List saved profiles
/cost-lean restore	Re-enable everything

CLI wrapper for applying profiles before a session:

npx cost-lean                     # show current context load
npx cost-lean --recommend         # generate profile recommendations
npx cost-lean --profile lean      # apply a saved profile
npx cost-lean --list              # list saved profiles
npx cost-lean --restore           # restore full config

# Then start your session
claude

Profiles disable unused MCPs (by filtering .mcp.json) and skills (by renaming SKILL.md → SKILL.md.disabled). A SessionEnd hook automatically restores your full config when the session ends.

/cost-optimize — Smart Recommendations

Analyzes spending patterns across 7 dimensions:

Analyzer	What It Detects
Model Tier	Opus sessions that could use Sonnet
Cache Efficiency	Low cache hit rates, short sessions wasting warmup
Context Bloat	Sessions exceeding 500K tokens
Project Patterns	Projects using Opus for routine work
Session Name Patterns	Task types correlated with model cost
Time Patterns	Expensive time-of-day patterns
Subagent Usage	Multi-model session overhead

/cost-dashboard — Interactive Visualization

Generates a self-contained HTML dashboard with Chart.js:

• KPI cards (total, today, 7-day, 30-day, cache efficiency)
• Daily spend trends, model tier doughnut, per-project bars
• Hour-of-day distribution, cumulative spend, sessions per day
• Searchable session table with cost tiers and summaries

/cost-budget — Spending Limits

/cost-budget set weekly 500     # Set $500/week limit
/cost-budget status             # Check current spend vs limits
/cost-budget clear              # Remove all limits

Alerts at 80% (yellow) and 100% (red) via the Stop hook.

Multi-Runtime Support

Context Lens auto-detects your runtime and applies the correct pricing:

Runtime	Pricing Formula	Config Path
Claude Code	Direct Anthropic pricing	~/.claude/.mcp.json
Cursor	(provider_price × 93%) + $0.25/1M tokens	~/.cursor/mcp.json

Detection is automatic via environment variables. All reports, audits, and profiles work with both runtimes.

Real-Time Cost Monitor

A Stop hook surfaces tips as system messages:

Threshold	Trigger	Tip
Context 200K	Peak context > 200K	Suggests /compact
Context 500K	Peak context > 500K	Warns about per-message cost
Cost $50	Session > $50	Suggests Sonnet if on Opus
Cost $200	Session > $200	Recommends new session
Opus routine	Every 20 Opus messages	Reminds about Sonnet
Low cache	Cache hit rate < 30%	Advises against clearing context

Each tip shows once per session.

How It Works

Session Start → Weekly digest check → Context snapshot
     ↓
Each turn → Cost monitor + budget check
     ↓
Session End → Session restore (if lean profile active)
           → Parse transcript → Classify → Calculate → Log
     ↓
On demand → /cost-tracker, /cost-audit, /cost-lean,
            /cost-optimize, /cost-dashboard, /cost-budget

Hooks

Event	Hook	Purpose
SessionStart	weekly-digest.js	Weekly comparison digest
SessionStart	context-snapshot.js	Log installed skills/MCPs for trending
Stop	cost-monitor.js	Context and cost threshold alerts
Stop	budget-check.js	Budget limit enforcement
SessionEnd	session-restore.js	Restore config if lean profile was active
SessionEnd	session-logger.js	Parse, classify, calculate, and log

Data Files

File	Purpose
~/.claude/cost-tracker/cost-log.jsonl	One JSON line per session
~/.claude/cost-tracker/context-snapshots.jsonl	Context load snapshots per session
~/.claude/cost-tracker/profiles/	Saved context profiles
~/.claude/cost-tracker/backups/	Config backups (auto-managed)
~/.claude/cost-tracker/budget.json	Budget limits
~/.claude/cost-tracker/dashboard.html	Last generated dashboard
~/.claude/cost-tracker/digests/	Weekly digest markdown files

All data stays local.

Supported Models & Pricing

Per million tokens (April 2026):

Model	Input	Output	Cache Write	Cache Read
Opus 4.7 / 4.6	$5.00	$25.00	$6.25	$0.50
Sonnet 4.6 / 4.5	$3.00	$15.00	$3.75	$0.30
Haiku 4.5	$1.00	$5.00	$1.25	$0.10

Pricing is centralized in lib/pricing-engine.js. Unknown models default to Sonnet with a pricing_estimated flag.

Architecture

plugins/cost-tracker/
  lib/                    # Shared modules
    pricing-engine.js     # Model pricing + Cursor support
    context-auditor.js    # 6 scanners for context bloat
    usage-correlator.js   # Skill/MCP usage tracking
    config-paths.js       # Runtime detection (CC vs Cursor)
    profile-manager.js    # Profile CRUD, backup/restore
  hooks/                  # Lifecycle hooks
  skills/                 # 6 slash commands
  bin/                    # CLI wrapper (cost-lean)
  evals/                  # Agent behavior tests
  __tests__/              # Jest test suites

Testing

npm install
npm test

218 tests across 13 suites covering pricing calculations, context scanning, usage correlation, profile management, transcript parsing, session classification, threshold logic, budget checks, digest generation, and report formatting.

Evals

Agent behavior tests validate that skills produce correct recommendations:

node plugins/cost-tracker/evals/run-evals.js

5 scenarios: vague cost help, skill bloat detection, Cursor pricing, MCP partial use, root cause analysis.

Platform Support

• Windows 11 (Git Bash) — includes path normalization
• macOS and Linux
• Cursor (auto-detected via environment variables)

Contributing

Issues and PRs welcome at github.com/ChewbaccaRoars/context-lens.

License

MIT