Your first eval starts here.

Install the CLI and try it

macOS or Linux. Requires uv and a running Docker daemon.

uv tool install https://github.com/portofcontext/agent-voyager-project/releases/download/cli-v0.7.0/avp_cli-0.7.0-py3-none-any.whl
avp hello-world

avp hello-world runs an example eval on three agents. LLM calls are mocked, so it needs no API key and costs nothing. avp eval share publishes it to a public URL, and avp update installs newer releases.

Connect your models

Save a key for each provider you use. Each command prompts for the key.

avp vault set openai

avp vault set anthropic

avp vault set google

Codex uses OpenAI; Claude Code uses Anthropic. Pi can use either, or Google. Install an agent with avp agent install codex, claude-code, or pi; avp agent list --builtin lists them all.

Start configuring agents

A Commission is one JSON file that configures a run: the model, the prompt, and what the agent may use. Save this as task.json, with an Anthropic key saved above.

{
  "avp_version": "0.1",
  "model": {
    "provider": "anthropic",
    "id": "claude-haiku-4-5",
    "credential": {"proxy": "anthropic"}
  },
  "prompt": "Say hello."
}

Install an agent and run the Commission on it.

avp agent install claude-code
avp run --agent claude-code --commission task.json

avp commission new --agent claude-code --model anthropic/claude-haiku-4-5 writes a Commission that lists the agent's tools, skills, subagents, and MCP servers, so you can turn each on or off. Compare Commissions and agents in an eval file. Runs on real models make paid calls.

Add your own harness (optional)

Any runtime can be an AVP agent: an image whose entrypoint answers ping, describe, and run. Write the adapter with the SDK, following the integration patterns, then install and check it.

avp agent install my-agent ./Dockerfile
avp conformance check --agent my-agent