Adapterly — AI Cost Analyzer

AI Cost Analyzer

Upload your OpenAI or Anthropic usage log — get a savings report in ~30 seconds.

How to get your logs

Three ways, from easiest to richest. The custom logger gives the best analysis because it includes prompt snippets — console exports only include token counts.

Export from OpenAI Usage dashboard

  1. Log in to platform.openai.com/usage
  2. Pick a date range (last 30 days is ideal — enough signal, not too much data)
  3. Click Export at the top right — downloads a CSV
  4. Upload that CSV above
What you get: per-day token counts and costs by model. Enough for the numeric substitution analysis. No prompt content, so the qualitative workload-type analysis will be skipped.

Export from Anthropic Console

  1. Log in to console.anthropic.com/settings/usage
  2. Pick a date range — last 30 days is a good default
  3. Click Export — downloads a CSV with input/output/cache token columns
  4. Upload that CSV above
What you get: per-day token counts by model, plus cache-read tokens (so the report can spot missed caching opportunities). No prompt content.

Drop-in logger — captures prompts too

One file, two lines of code, runs alongside your production traffic. Writes to ~/.adapterly/logs/YYYY-MM-DD.jsonl locally. Nothing is uploaded until you decide to.

1. Download the logger:

↓ adapterly_logger.py  or via curl:

# save it next to your app code
curl -O https://adapterly.ai/tools/cost-analyzer/adapterly_logger.py

2. Add two lines to your app startup:

from adapterly_logger import instrument
instrument()

That's it. Every subsequent OpenAI or Anthropic call is logged. Your app code doesn't change.

3. Run your app normally for a few days. A week of traffic gives a solid analysis; longer is better.

4. Concatenate the daily files and upload:

cat ~/.adapterly/logs/*.jsonl > my-usage.jsonl
# then upload my-usage.jsonl above
Privacy: Only the first 400 chars of the first user message is captured as a snippet (for workload classification). Full prompts and responses are never logged. Set ADAPTERLY_LOGGER_NO_PROMPTS=1 to skip snippets entirely.
Why this gives the best analysis: the qualitative panel (workload types, prompt patterns, prioritized recommendations) requires prompt samples. Console CSV exports don't include them — only the logger does.

Add as an MCP tool in your AI editor

Analyze usage directly in Claude Desktop, Cursor, VS Code Copilot Chat, Windsurf, or ChatGPT — no file upload, no browser. Ask "analyze my usage log", "how much would I save moving GPT-4o to open-source?", or "list all EU-hostable model options".

Endpoint:

https://adapterly.ai/tools/cost-analyzer/mcp/

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "adapterly-cost-analyzer": {
      "url": "https://adapterly.ai/tools/cost-analyzer/mcp/"
    }
  }
}

Cursor — edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "adapterly-cost-analyzer": {
      "url": "https://adapterly.ai/tools/cost-analyzer/mcp/"
    }
  }
}

VS Code Copilot Chat — create .vscode/mcp.json in your workspace:

{
  "servers": {
    "adapterly-cost-analyzer": {
      "type": "http",
      "url": "https://adapterly.ai/tools/cost-analyzer/mcp/"
    }
  }
}

ChatGPT — not yet supported. ChatGPT requires MCP servers to implement OAuth 2.0 authentication; this server currently runs open (no auth) because Claude Desktop, Cursor and VS Code don't require it. OAuth support will be added when there's demand — until then, ChatGPT users can either use the web upload above or switch to one of the clients above.

Available tools: analyze_usage_log, list_supported_models, estimate_savings, explain_substitution. Plus adapterly://pricing resource for the live pricing table.
Why MCP: your usage log stays in your machine when using local file paths through the client; only the numbers we compute travel over the wire. No file upload UI, no browser tab — just ask a question.