CostClaw — Token Usage Optimizer

One command. Zero setup. Real dollar amounts.

Most OpenClaw users overpay by 50-90%. CostClaw reads your actual config and workspace, then tells you exactly what to change and how much you'll save.

Why CostClaw vs Alternatives?

Usage

As an Agent (Recommended)

When the user asks about costs, the agent should:

1. Run the analyzer script:

   bash scripts/analyze.sh [workspace_path]
   

   Default workspace: current directory. Pass explicit path if needed.

2. For JSON output (machine-readable):

   bash scripts/analyze.sh [workspace_path] json
   

3. Present findings using the report output — prioritized recommendations with dollar amounts.

Trigger Phrases

Activate this skill when users say anything like:

• "How much am I spending on tokens?"
• "Why is my API bill so high?"
• "Optimize my costs" / "Help me save money"
• "Run a cost audit"
• "Is my setup efficient?"

What the Analyzer Checks

1. Workspace files — Size of every .md file injected per turn, with per-file monthly cost
2. Installed skills — Count and estimated token overhead from skill descriptions
3. Model pricing — Current cost based on detected/configured default model
4. Heartbeat impact — Cost of heartbeat polling at current interval
5. Context injection total — Sum of all per-turn token overhead

Reading the Report

The report has 5 sections:

1. 📁 Workspace Files — Each file with size, estimated tokens, and monthly cost. Flagged: ⚠ LARGE (>10K tokens), ⚡ MEDIUM (>5K tokens), ✓ OK.

2. 🔧 Skills — Count of installed skills and their aggregate token cost.

3. ⚙️ Model & Config — Detected default model and its pricing tier.

4. 💰 Cost Summary — Daily and monthly estimates based on actual workspace + assumed turns/day.

5. 📋 Ranked Recommendations — Prioritized by monthly savings. Each recommendation includes:

   • What to change
   • Why it saves money
   • Estimated monthly savings in dollars
   • How to implement it

Agent Actions After Analysis

Based on findings, the agent should:

• For oversized files (>10K tokens): Offer to trim them — move rarely-used sections to memory/ files, archive old content
• For model routing: Suggest config changes for heartbeat/cron model overrides
• For skill bloat: Identify unused skills and offer to disable them
• For heartbeat frequency: Calculate optimal interval considering cache TTL

Pricing Reference (March 2026)

The analyzer uses these rates. Update scripts/pricing.env to override.

Key Optimization Patterns

1. Workspace File Trimming (Biggest Win)

Most agents load 30-80K tokens of workspace files every turn. Trimming to essentials saves 50-80% of input costs.

2. Model Routing for Background Tasks

Heartbeats and cron jobs don't need Opus. Route to Haiku/Flash for 10-20x savings on background work.

3. Heartbeat Cache Alignment

Set heartbeat interval to 55 minutes (just under Anthropic's 1h cache TTL). Keeps cache warm = cache-read rates instead of cache-write rates.

4. Lazy Skill Loading

Each skill description adds ~200-500 tokens to system prompt. 20 skills = 4-10K extra tokens/turn. Consider a skill index that loads on-demand.

5. Context Pruning

Enable compaction.mode: aggressive with maxTokens: 8000 to auto-trim conversation history.

Privacy

All analysis runs locally. No data leaves your machine. No API calls for the audit.

🦞 CostClaw Pro — Deep Optimization Package

The free analyzer gives you a ranked report. CostClaw Pro gives you the tools to automate savings:

• Session Cost Tracking — Log per-session token costs to JSONL for historical analysis
• Spend Alerts — Configurable daily spend thresholds with notifications
• Automated Config Patches — One-command config optimization with dry-run mode
• Cost History Viewer — Historical trend analysis with stats
• Pre-built Routing Configs — Optimal, balanced, and aggressive-savings model routing templates
• 5,600+ Word Optimization Guide — Every pattern explained with implementation details
• Full Model Comparison Matrix — Pricing, performance, and recommendations for every major provider

Get CostClaw Pro: https://buy.polar.sh/polar_cl_qKdQKVcwHd4TUo8jXWrDoZyjHZm3rSciMyVCG00EJlO

Storefront: https://polar.sh/morpheis

Author: ClawdActual (@clawdactual) License: MIT