AgentProbe

Test how well AI agents interact with your CLI tools. AgentProbe runs Claude Code against any command-line tool and provides actionable insights to improve Agent Experience (AX) - helping CLI developers make their tools more AI-friendly.

Quick Start

# No installation needed - run directly with uvx
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test vercel --scenario deploy

# Or install locally for development
uv sync
uv run agentprobe test git --scenario status

Authentication

AgentProbe supports multiple authentication methods to avoid environment pollution:

Get an OAuth Token

First, obtain your OAuth token using Claude Code:

claude setup-token

This will guide you through the OAuth flow and provide a token for authentication.

Method 1: Token File (Recommended)

# Store token in a file (replace with your actual token from claude setup-token)
echo "your-oauth-token-here" > ~/.agentprobe-token

# Use with commands
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test vercel --scenario deploy --oauth-token-file ~/.agentprobe-token

Method 2: Config Files

Create a config file in one of these locations (checked in priority order):

# Global user config (replace with your actual token from claude setup-token)
mkdir -p ~/.agentprobe
echo "your-oauth-token-here" > ~/.agentprobe/config

# Project-specific config (add to .gitignore)
echo "your-oauth-token-here" > .agentprobe
echo ".agentprobe" >> .gitignore

# Then run normally without additional flags
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test vercel --scenario deploy

Method 3: Environment Variables (Legacy)

# Replace with your actual token from claude setup-token
export CLAUDE_CODE_OAUTH_TOKEN="your-oauth-token-here"
# Note: This may affect other Claude CLI processes

Recommendation: Use token files or config files for better process isolation.

What It Does

AgentProbe launches Claude Code to test CLI tools and provides Agent Experience (AX) insights on:

• AX Score (A-F) based on turn count and success rate
• CLI Friction Points - specific issues that confuse agents
• Actionable Improvements - concrete changes to reduce agent friction
• Real-time Progress - see agent progress with live turn counts

Community Benchmark

Help us build a comprehensive benchmark of CLI tools! The table below shows how well Claude Code handles various CLIs.

View detailed results →

Commands

Test Individual Scenarios

# Test a specific scenario (with uvx)
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test gh --scenario create-pr

# With authentication token file
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test gh --scenario create-pr --oauth-token-file ~/.agentprobe-token

# Test multiple runs for consistency analysis
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test vercel --scenario deploy --runs 5

# With custom working directory
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test docker --scenario run-nginx --work-dir /path/to/project

# Show detailed trace with message debugging (disables progress indicators)
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test gh --scenario create-pr --verbose

# ⚠️ DANGEROUS: Run without permission prompts (use only in safe environments)
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test docker --scenario run-nginx --yolo

Benchmark Tools

# Test all scenarios for one tool
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe benchmark vercel

# Test all scenarios with authentication
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe benchmark vercel --oauth-token-file ~/.agentprobe-token

# Test all available tools and scenarios
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe benchmark --all

# ⚠️ DANGEROUS: Run all benchmarks without permission prompts
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe benchmark --all --yolo

Reports

# Generate reports (future feature)
uv run agentprobe report --format markdown --output results.md

Debugging and Verbose Output

The --verbose flag provides detailed insights into how Claude Code interacts with your CLI:

# Show full message trace with object types and attributes
uvx --from git+https://github.com/nibzard/agentprobe.git agentprobe test gh --scenario create-pr --verbose

Verbose output includes:

• Message object types (SystemMessage, AssistantMessage, UserMessage, ResultMessage)
• Message content and tool usage
• SDK object attributes and debugging information
• Full conversation trace between Claude and your CLI

⚠️ YOLO Mode (Use with Extreme Caution)

The --yolo flag enables autonomous execution without permission prompts, allowing Claude to run ANY command without user approval:

# WARNING: Only use in isolated, safe environments
agentprobe test docker --scenario build-app --yolo

Security Considerations:

• ONLY use in containerized or sandboxed environments
• Claude can execute arbitrary commands including rm -rf, network calls, system modifications
• No safety guardrails - Claude has full system access
• Intended for CI/CD pipelines, testing environments, or research purposes
• NEVER use on production systems or with sensitive data

This mode is equivalent to running Claude Code with --dangerously-skip-permissions.

Example Output

Single Run (Default)

⠋ Agent running... (Turn 3, 12s)
╭───────────────────────────── AgentProbe Results ─────────────────────────────╮
│ Tool: vercel | Scenario: deploy                                               │
│ AX Score: B (12 turns, 80% success rate)                                      │
│                                                                               │
│ Agent Experience Summary:                                                     │
│ Agent completed deployment but needed extra turns due to unclear progress     │
│ feedback and ambiguous success indicators.                                    │
│                                                                               │
│ CLI Friction Points:                                                          │
│ • No progress feedback during build process                                   │
│ • Deployment URL returned before actual completion                            │
│ • Success status ambiguous ("building" vs "deployed")                        │
│                                                                               │
│ Top Improvements for CLI:                                                     │
│ 1. Add --status flag to check deployment progress                             │
│ 2. Include completion status in deployment output                             │
│ 3. Provide structured --json output for programmatic usage                    │
│                                                                               │
│ Expected turns: 5-8 | Duration: 23.4s | Cost: $0.012                         │
│                                                                               │
│ Use --verbose for full trace analysis                                         │
╰───────────────────────────────────────────────────────────────────────────────╯

Multiple Runs (Aggregate)

╭──────────────────────── AgentProbe Aggregate Results ────────────────────────╮
│ Tool: vercel | Scenario: deploy                                               │
│ AX Score: C (14.2 avg turns, 60% success rate) | Runs: 5                      │
│                                                                               │
│ Consistency Analysis:                                                         │
│ • Turn variance: 8-22 turns                                                   │
│ • Success consistency: 60% of runs succeeded                                  │
│ • Agent confusion points: 18 total friction events                            │
│                                                                               │
│ Consistent CLI Friction Points:                                               │
│ • Permission errors lack clear remediation steps                              │
│ • No progress feedback during deployment                                      │
│ • Build failures don't suggest next steps                                     │
│                                                                               │
│ Priority Improvements for CLI:                                                │
│ 1. Add deployment status polling with vercel status                           │
│ 2. Include troubleshooting hints in error messages                            │
│ 3. Provide progress indicators during long operations                          │
│                                                                               │
│ Avg duration: 45.2s | Total cost: $0.156                                      │
╰───────────────────────────────────────────────────────────────────────────────╯

Contributing Scenarios

We welcome scenario contributions! Help us test more CLI tools:

1. Fork this repository
2. Add your scenarios under scenarios/<tool-name>/
3. Run the tests and update the benchmark table
4. Submit a PR with your results

Scenario Format

Simple Text Format

Create simple text files with clear prompts:

# scenarios/stripe/create-customer.txt
Create a new Stripe customer with email test@example.com and
add a test credit card. Return the customer ID.

Enhanced YAML Format (Recommended)

Use YAML frontmatter for better control and metadata:

# scenarios/vercel/deploy-complex.txt
---
version: 2
created: 2025-01-22
tool: vercel
permission_mode: acceptEdits
allowed_tools: [Read, Write, Bash]
model: opus
max_turns: 15
complexity: complex
expected_turns: 8-12
description: "Production deployment with environment setup"
---
Deploy this Next.js application to production using Vercel CLI.
Configure production environment variables and ensure the deployment
is successful with proper domain configuration.

YAML Frontmatter Options:

• model: Override default model (sonnet, opus)
• max_turns: Limit agent interactions
• permission_mode: Set permissions (acceptEdits, default, plan, bypassPermissions)
• allowed_tools: Specify tools ([Read, Write, Bash])
• expected_turns: Range for AX scoring comparison
• complexity: Scenario difficulty (simple, medium, complex)

Running Benchmark Tests

# Test all scenarios for a tool
uv run agentprobe benchmark vercel

# Test all tools
uv run agentprobe benchmark --all

# Generate report (placeholder)
uv run agentprobe report --format markdown

Architecture

AgentProbe follows a simple 4-component architecture:

1. CLI Layer (cli.py) - Typer-based command interface with progress indicators
2. Runner (runner.py) - Executes scenarios via Claude Code SDK with YAML frontmatter support
3. Analyzer (analyzer.py) - AI-powered analysis using Claude to identify friction points
4. Reporter (reporter.py) - AX-focused output for CLI developers

Agent Experience (AX) Analysis

AgentProbe uses Claude itself to analyze agent interactions, providing:

• Intelligent Analysis: Claude analyzes execution traces to identify specific friction points
• AX Scoring: Automatic scoring based on turn efficiency and success patterns
• Contextual Recommendations: Actionable improvements tailored to each CLI tool
• Consistency Tracking: Multi-run analysis to identify systematic issues

This approach avoids hardcoded patterns and provides nuanced, tool-specific insights that help CLI developers understand exactly where their tools create friction for AI agents.

Prompt Management & Versioning

AgentProbe uses externalized Jinja2 templates for analysis prompts:

• Template-based Prompts: Analysis prompts are stored in prompts/analysis.jinja2 for easy editing and iteration
• Version Tracking: Each analysis includes prompt version metadata for reproducible results
• Dynamic Variables: Templates support contextual variables (scenario, tool, trace data)
• Historical Comparison: Version tracking enables comparing results across prompt iterations

# Prompt templates are automatically loaded from prompts/ directory
# Version information is tracked in prompts/metadata.json
# Analysis results include prompt_version field for tracking

Requirements

• Python 3.10+
• uv package manager
• Claude Code SDK (automatically installed)

Key Features

🎯 Agent Experience (AX) Focus

• AX Scores (A-F) based on turn efficiency and success rate
• Friction Point Analysis identifies specific CLI pain points
• Actionable Recommendations for CLI developers

📊 Progress & Feedback

• Real-time Progress with live turn count and elapsed time
• Consistency Analysis across multiple runs
• Expected vs Actual turn comparison using YAML metadata

🔧 Advanced Scenario Control

• YAML Frontmatter for model selection, permissions, turn limits
• Multiple Authentication methods with process isolation
• Flexible Tool Configuration per scenario

Available Scenarios

Current test scenarios included:

• GitHub CLI (gh/)
  • create-pr.txt - Create pull requests
• Vercel (vercel/)
  • deploy.txt - Deploy applications to production
  • preview-deploy.txt - Deploy to preview environment
  • init-project.txt - Initialize new project with template
  • env-setup.txt - Configure environment variables
  • list-deployments.txt - List recent deployments
  • domain-setup.txt - Add custom domain configuration
  • rollback.txt - Rollback to previous deployment
  • logs.txt - View deployment logs
  • build-local.txt - Build project locally
  • ax-test.txt - Simple version check (AX demo)
  • yaml-options-test.txt - YAML frontmatter demo
• Docker (docker/)
  • run-nginx.txt - Run nginx containers
• Wrangler (Cloudflare) (wrangler/)
  • Multiple deployment and development scenarios

Browse all scenarios →

Development

# Install with dev dependencies
uv sync --extra dev

# Format code
uv run black src/

# Lint code
uv run ruff check src/

# Run tests (when implemented)
uv run pytest

See TASKS.md for the development roadmap and task tracking.

Programmatic Usage

import asyncio
from agentprobe import test_cli

async def main():
    result = await test_cli("gh", "create-pr")
    print(f"Success: {result['success']}")
    print(f"Duration: {result['duration_seconds']}s")
    print(f"Cost: ${result['cost_usd']:.3f}")

asyncio.run(main())

License

MIT