Your AI agents are re-buying the same tokens every turn.

Agent Cost Lens reads your agents' usage metadata, shows where the API money actually goes, and simulates what proper prompt caching would recover — up to $1,918 (simulated) on our public demo workload alone.

Get your key — $29/month
Stripe checkout → your API key on the next page → analyzing in under five minutes. Cancel anytime; your key deactivates automatically.

How it works

  1. Sweep locally. The MCP client reads your Claude Code transcripts or your app's API logs on your machine.
  2. Scrub locally. Repo and session names become salted pseudonyms before anything leaves. Prompt text and code are never collected — the wire format has no field for them.
  3. Analyze hosted. Usage metadata (token counts, model IDs, dates) goes to the analyzer; you get spend by model/day/source, your real cache-hit rate, and what a proper caching setup would recover — every estimate labeled simulated and up to, because honest math beats a sales curve.

Audit us before you trust us

The client ships a preview_upload tool that prints the exact payload that would be sent — before anything is sent, no key required. If you can find a prompt, a file path, or a real project name in it, don't buy.

Quickstart

curl -LO https://lens.r-lattice.com/download
pip install agent_cost_lens_mcp-0.1.0.tar.gz
# add to your MCP config (Claude Code, Claude Desktop, or any MCP host):
{
  "mcpServers": {
    "agent-cost-lens": {
      "command": "acl-mcp",
      "env": {
        "LENS_SERVER_URL": "https://lens.r-lattice.com",
        "LENS_API_KEY": "acl_..."   // from your claim page after checkout
      }
    }
  }
}

Then ask your agent to run preview_upload (the audit) and analyze_costs (the answer).

Who this is for

Teams and solo builders running Claude / Anthropic API agents — custom agent loops, Claude Code fleets, LangChain/raw-SDK apps — who suspect their context re-sends are quietly costing real money. If your cache-hit rate is already 95%+, we'll tell you that too: the report says "caching can't help this workload" when it's true.