What we see, and what we never see

What we see

• Your agent's answers to our questions, after the runner redacts emails, phone numbers, keys, tokens and file paths. We redact again on arrival, before anything is stored.
• Small exam files your agent was asked to create inside the runner's sandbox folder (for example, a one-line text file), capped at 2 KB each.
• Setup counts: how many memory files, skills and MCP servers were found. The runner also shows and sends their names (for example "AGENTS.md" or a skill folder name) unless you pass --counts-only. We use the counts for your prescription and don't store the names.
• An optional label you choose for the report.

What we never see

• The contents of your memory files, skills, notes or code.
• Connector settings: commands, URLs, environment variables, tokens.
• Your agent's raw output (it stays in .teachmyagent-exam/ on your machine), your folder paths, or anything outside the sandbox.
• Your account, email or name: sessions are anonymous.

Storage

• Sessions are anonymous. Your IP address is used only for in-memory rate limiting and is never written to disk.
• Answers are deleted as soon as they're graded; only the report card is kept.
• Report cards are auto-deleted after 30 days unless you choose to share them (--share, or share: true on finish_exam).
• Question content is fictional, and any "secrets" in it are fake.

Verify it yourself

• The runner is open source (MIT): read every file.
• --dry-run runs fictional sample questions, prints every path it would check and the exact payload, and makes zero network calls.
• A built-in network guard logs every connection the runner process attempts, and blocks everything except our server (and everything at all in a dry run). Each run saves netlog.json and payload.json next to the report.
• Your agent is a separate program and uses its own network (its model provider, its connectors) as it always does. Watch it with your OS firewall or lsof -i if you want to.

Online exam: clean-room setup

For the online (MCP) exam, test in a throwaway environment so nothing personal is within reach:

1. Create an empty folder (mkdir ~/agent-exam && cd ~/agent-exam) and start your agent there.
2. Claude Code: run claude --strict-mcp-config --mcp-config exam.json, where exam.json lists only TeachMyAgent.
3. Codex: use a separate CODEX_HOME whose config.toml lists only TeachMyAgent.
4. Gemini CLI: put a .gemini/settings.json in the empty folder with only TeachMyAgent.
5. Grok Bot: start a new chat and switch off personal connectors (mail, calendar, drive) for the run.
6. Don't paste personal data into answers. We redact common patterns anyway.

These setup flags change between versions. Check your agent's docs if one doesn't work.