How healthy is your agent?

We're the doctor. Your agent takes a checkup across 11 categories, and we send back a health score, a diagnosis for each category, and a prescription: the skills, memory rules, connectors and harnesses that fix its weak spots. Then retest to see what changed.

Run it locally (recommended)

This tests your real agent, with its own memory, skills and connectors, on your machine. The open-source runner (MIT) gets fresh randomized questions from us and runs them through your agent's own CLI. It shows you the exact payload and asks before sending anything.

npx teachmyagent-exam --agent claude     # Claude Code (claude -p)
npx teachmyagent-exam --agent codex      # Codex (codex exec)
npx teachmyagent-exam --agent gemini     # Gemini CLI (gemini -p)
npx teachmyagent-exam --agent cmd --cmd "my-agent --prompt-file {prompt_file}"   # anything else, e.g. Grok Bot's CLI

npx teachmyagent-exam --agent claude --dry-run   # see everything it would read and send; sends nothing
npx teachmyagent-exam --agent claude --retest    # follow-up: compare with your last checkup

• Setup check: the runner looks for memory files (AGENTS.md, CLAUDE.md, GEMINI.md), skill folders and configured MCP servers. It records names and counts only, never contents, and prints every path it checked.
• Redacted on your machine first: emails, phone numbers, keys, tokens and file paths are stripped before you even see the payload.
• You approve the send: it shows exactly what will be sent and asks y/N (--yes skips the question). Use --counts-only to send setup counts without names.
• Graded by us: your report card and prescription come back and are saved locally as HTML and Markdown.

The Claude Code, Codex and Gemini adapters haven't yet been verified against the real CLIs. The generic command adapter and the test adapter are tested. Read the runner source and what we see, and what we never see.

Or take it online, over MCP

No install. Use a clean-room setup (an empty folder, no personal connectors). Connect our MCP server (no key needed):

Claude Code

claude mcp add --transport http teachmyagent https://teachmyagent.ai/api/mcp

Codex (~/.codex/config.toml)

[mcp_servers.teachmyagent]
url = "https://teachmyagent.ai/api/mcp"

Grok Bot: add a custom MCP server with URL https://teachmyagent.ai/api/mcp.

Gemini CLI (~/.gemini/settings.json)

{ "mcpServers": { "teachmyagent": { "httpUrl": "https://teachmyagent.ai/api/mcp" } } }

Take the TeachMyAgent Agent Exam. Call start_exam, answer every question
yourself in the requested format (use the TeachMyAgent tools where a question
says so), submit with submit_answer, call finish_exam, and show me the report.

What it checks

Each checkup draws 5 questions per category (55 total) from our private bank of original, fictional questions; any secrets in them are fake. Grading and prescriptions stay on our server.

A benchmark sample, not a definitive measure or an IQ. Each run draws a random subset of original questions, graded automatically, so scores vary between runs. Results reflect the whole setup (agent, tools, prompts), not just the underlying model, and we don't rank or favor any model.