claude-dentist

7 out of 10 Claude Dentists agree this package is GREAT and not COULD BE BETTER.

Agent Experience (AX) testing for tools built for Claude Code. Asserts X out of Y agents report your tool works GREAT and not COULD BE BETTER.

Agents run with bypassPermissions enabled and can execute shell commands, read/write files, and install packages without confirmation. Always run in a sandboxed environment.

Usage

import pytest
from claude_dentist import claude_dentist

@pytest.mark.anyio
async def test_agent_experience() -> None:
    await claude_dentist(
        runs=3,
        min_passes=2,
        prompt="Assess my-tool by using it. Install with `uv add my-tool`.",
        max_turns=15,
        deadline_seconds=60,
    )

Self-assessment

This is the test used to assess claude-dentist itself:

@pytest.mark.agent_experience
@pytest.mark.anyio
async def test_claude_can_use_claude_dentist() -> None:
    await claude_dentist(
        runs=10,
        min_passes=9,
        prompt=textwrap.dedent(f"""\
            Assess the claude-dentist library. Create a temporary project
            in /tmp, install it with
            `uv add "claude-dentist @ {CLAUDE_DENTIST_ROOT}"`, and try
            it out. Use a maximum of 3 runs to save on costs.
        """),
        options=ClaudeAgentOptions(max_turns=30),
        deadline_seconds=600,
    )

Dissenting opinions

Three Claude Dentists reported this library COULD BE BETTER:

• Dentist #2 found that sub-agents kept hitting the max turns limit before completing, causing ERROR verdicts.
• Dentist #8 discovered that verdict parsing is fragile — an agent that mentions "COULD BE BETTER" anywhere in its narrative gets misclassified, even when the final judgment is GREAT.
• Dentist #10 actually said GREAT, but fell victim to the same verdict parsing bug that Dentist #8 reported.