Agentops

tests

An MCP server that exposes AI-coding-agent operations analytics as tools any MCP client can call. Ask your assistant "how much did I spend on Claude Code this week?", "what dangerous commands did my agent attempt?", or "where are my tokens wasted?" — answered from the transcripts Claude Code already writes locally. Read-only, no telemetry.

Tools

Tool	Returns
cost_summary	tokens, cache hit rate, est. dollars, top projects & tools
audit_summary	commands run, file writes, errors, sensitive-path access
safety_report	dangerous actions the agent attempted (force-push, rm -rf, curl|sh)
efficiency_report	re-read waste, cache misses, concrete fixes

Each takes days (default 30) and an optional project filter.

Install

pipx install git+https://github.com/Ludoonus/agentops-mcp
# or: uvx --from git+https://github.com/Ludoonus/agentops-mcp agentops-mcp

Use with Claude Desktop (or any MCP client)

Add to your MCP config:

{
  "mcpServers": {
    "agentops": { "command": "agentops-mcp" }
  }
}

Then ask your assistant about your Claude Code cost, audit trail, safety, or efficiency.

How it works

Reads ~/.claude/projects/*.jsonl (the transcripts Claude Code writes), parses the API-reported usage and tool calls, and serves the analysis over MCP (stdio). Nothing leaves your machine; the server only reads.

Companion projects

• Operator — the same analytics as a CLI/TUI console.
• cc-powerpack — guardrail hooks that prevent the dangerous actions this server reports.

Go deeper

These tools are the practical layer; The Claude Code Operator's Handbook is the full playbook — 18 chapters on running AI coding agents safely and efficiently (threat model, guardrails, cost, workflows, recovery, scaling). $29, with a free 3-chapter sample.

License

MIT.