privacy
We collect performance signals, not transcripts.
The useful product data is aggregate workflow intelligence: token buckets, waste patterns, known public tooling fingerprints, parser status, and redaction counts.
Agent Analyzer is an independent third-party tool and is not affiliated with Anthropic, OpenAI, Anysphere, Amazon, or Google. The public flow is local-first: raw supported agent logs stay on your machine, and only the sanitized report JSON is uploaded after review.
What We Receive
- Sanitized report JSON generated by the local CLI.
- Numeric and categorical metrics such as turns, estimated tokens, rereads, retry depth, and score bucket.
- Known public ecosystem IDs from the allowlisted registry.
- Counts of unknown private MCPs, skills, plugins, and tools, without their names.
What We Do Not Receive
- Raw agent logs in the public launch flow.
- Raw SQLite state stores and local session stores. Cursor, Kiro, Google Antigravity, and bounded Codex diagnostic extraction is read-only, limited to known rows, and analyzed locally without writing database snapshots. GitHub Copilot extraction reads Copilot CLI session/log files and VS Code Copilot Chat session JSON locally.
- Prompt text, tool output, file contents, command arguments, secret values, repo names, usernames, hostnames, or full file paths.
- Unknown private tool names unless a future explicit opt-in path is added.
Retention
Report pages are retained as private-link artifacts so users can return to them later. Aggregate analytics may be retained when they match the allowlist above. Raw logs are not retained because the public flow does not upload them.
run locally
Generate the sanitized report yourself.
npx --yes agent-analyzer@latest runAnalyzes supported agent logs on your machine, asks before upload, and uses the sanitized report to build a custom plugin for the waste it finds.