AgentEval.Cli 0.23.0-beta

AgentEval CLI

Command-line interface for AgentEval — the comprehensive .NET evaluation toolkit for AI agents. Evaluate any OpenAI-compatible agent, run compliance and security benchmark suites, manage the canonical .agenteval/ output store, and serve the Mission Control web portal — all from the terminal or your CI/CD pipeline.

Installation

dotnet tool install --global AgentEval.Cli --prerelease

Compatibility

AgentEval.Cli ships in lockstep with the AgentEval library — both are released together at the same version from one pipeline. The pinned Microsoft Agent Framework / Microsoft.Extensions.AI versions for each release are recorded in the CHANGELOG.

The tool multi-targets net8.0 and net10.0; the dotnet-tool installer picks the highest compatible runtime. The mc serve portal requires .NET 10.

Quick Start

Scaffold a test dataset

agenteval init
agenteval init -o my-tests.yaml
agenteval init --format json

Run evaluations

# Against Azure OpenAI
agenteval eval --azure --endpoint https://myresource.openai.azure.com/ --deployment-name gpt-4o --dataset agenteval.yaml

# Against OpenAI directly
agenteval eval --endpoint https://api.openai.com/v1 --model gpt-4o --dataset agenteval.yaml

# Against a local Ollama model
agenteval eval --endpoint http://localhost:11434/v1 --model llama3 --dataset agenteval.yaml

Stochastic evaluation (multi-run)

agenteval eval --azure --endpoint https://myresource.openai.azure.com/ --deployment-name gpt-4o --dataset agenteval.yaml --runs 5 --success-threshold 0.9

Run benchmark families

# List every registered family with its presets and cost tiers
agenteval bench --list

# Compliance: GDPR / EU AI Act (audit-chain evidence written to .agenteval/)
agenteval bench gdpr --subject my-agent --preset standard --response-file answer.txt
agenteval bench eu-ai-act --subject my-agent --input "..." --response-file answer.txt

# Security: OWASP LLM Top 10 / MITRE ATLAS / NIST AI RMF
agenteval bench owasp --subject my-agent
agenteval bench mitre --subject my-agent
agenteval bench nist  --subject my-agent

# Quality & performance
agenteval bench agentic --subject my-agent
agenteval bench perf latency --subject my-agent --azure-from-env
agenteval bench longmemeval --subject my-agent --preset subset
agenteval bench memory --subject my-agent --preset quick

Red team security scanning

# Run all attack types
agenteval redteam --azure --endpoint https://myresource.openai.azure.com/ --deployment-name gpt-4o --intensity moderate

# Run specific attacks, export SARIF for CI
agenteval redteam --azure --endpoint https://myresource.openai.azure.com/ --deployment-name gpt-4o --attacks PromptInjection,Jailbreak --format sarif

Manage the workspace

agenteval init-workspace --name "My Solution"   # bootstrap .agenteval/
agenteval doctor                                # validate structure + content hashes
agenteval migrate                               # dry-run legacy → canonical layout
agenteval migrate --apply                       # apply the migration

Render reports (no LLM cost)

# PDF from existing compliance evidence
agenteval compliance render --regulation gdpr --subject my-agent

# Markdown from existing agentic benchmark results
agenteval render --benchmark agentic --subject my-agent

Mission Control web portal (requires .NET 10)

agenteval mc serve --workspace .            # GraphQL + REST + SPA on one port (default 5000)
agenteval mc doctor                         # verify the portal bundle is intact

List metrics, attacks, exporters, and datasets

agenteval list
agenteval list --type metrics
agenteval list --type attacks

Authentication

The eval and redteam commands support two endpoint modes: Azure OpenAI (--azure) and OpenAI-compatible (--endpoint).

Azure OpenAI (--azure)

The --azure flag uses AzureOpenAIClient. Both --endpoint and --deployment-name are required:

# Key from env var
export AZURE_OPENAI_API_KEY=sk-...
agenteval eval --azure --endpoint https://myresource.openai.azure.com/ --deployment-name gpt-4o --dataset agenteval.yaml

Note: --deployment-name is the name you gave your model deployment in Azure AI Foundry, not the underlying model name.

OpenAI-compatible (--endpoint)

For OpenAI, Ollama, Groq, vLLM, LM Studio, Together.ai, or any OpenAI-compatible API:

# OpenAI (set OPENAI_API_KEY or use --api-key)
agenteval eval --endpoint https://api.openai.com/v1 --model gpt-4o --dataset agenteval.yaml --api-key sk-...

# Local Ollama (no key needed)
agenteval eval --endpoint http://localhost:11434/v1 --model llama3 --dataset agenteval.yaml

Commands

Requirements

• .NET 8.0 or .NET 10.0 (mc serve requires .NET 10.0)
• An AI agent endpoint (Azure OpenAI, OpenAI, Ollama, or any OpenAI-compatible API) for eval / redteam / LLM-graded benchmarks
• Built on Microsoft Agent Framework (MAF) and Microsoft.Extensions.AI

Documentation

• AgentEval Documentation
• Changelog

Contributing

Contributions are welcome! Please open an issue or pull request on the main repository.

License

MIT License. See LICENSE for details.