CLI Reference

All commands work from your project root (where .agentbreak/ lives).

`agentbreak init`

Create .agentbreak/ directory with default application.yaml and scenarios.yaml. Auto-detects your provider (OpenAI/Anthropic), framework, and MCP usage. Generates scenarios.yaml with a standard preset based on what's enabled.

agentbreak init

`agentbreak serve`

Start the chaos proxy.

agentbreak serve              # start on default port (5005)
agentbreak serve -v           # verbose logging
agentbreak serve --label "description"   # label for run history

`agentbreak validate`

Check your config files for errors.

agentbreak validate
agentbreak validate --test-connection   # also test upstream auth (proxy mode)

The --test-connection flag makes a lightweight request to each proxy-mode upstream to verify connectivity and auth. Results:

OK — upstream reachable, auth works
AUTH FAILED (401) — bad API key or token
FORBIDDEN (403) — key valid but insufficient permissions
CONNECTION FAILED — wrong URL or upstream is down
TIMEOUT — upstream too slow to respond

`agentbreak inspect`

Connect to your MCP server and discover available tools, resources, and prompts. Writes .agentbreak/registry.json.

agentbreak inspect

`agentbreak verify`

Run the test suite (pytest).

agentbreak verify

`agentbreak mcp-server`

Start AgentBreak as an MCP server (for advanced use). Most users should use the Claude Code plugin instead.

agentbreak mcp-server    # requires: pip install mcp

`agentbreak history`

List past runs (requires history.enabled: true in application.yaml).

agentbreak history
agentbreak history show <id>
agentbreak history compare <a> <b>

Endpoints

When the proxy is running:

Endpoint	Description
`POST /v1/chat/completions`	OpenAI proxy
`POST /v1/messages`	Anthropic proxy
`POST /mcp`	MCP proxy
`GET /_agentbreak/scorecard`	LLM scorecard
`GET /_agentbreak/mcp-scorecard`	MCP scorecard
`GET /_agentbreak/requests`	Recent LLM requests
`GET /_agentbreak/mcp-requests`	Recent MCP requests
`POST /_agentbreak/reset`	Reset all runtime counters
`GET /_agentbreak/history`	List past runs (if history enabled)
`GET /_agentbreak/history/{id}`	Get details for a specific run
`GET /healthz`	Health check