Quick Start

Bootstrap Your Environment

Detect your platform, probe for existing backends, and generate a models.json:

az bootstrap

This checks common ports (11434 for Ollama, 8080 for llama.cpp, 8000 for vLLM), recommends the best backend for your platform, and optionally installs it. Use --non-interactive for CI or --skip-model to skip model downloads.

Initialize a Project

cd your-project
az init --private

This creates .agentzero/ with:

• policy.yml — security policy rules
• settings.toml — default provider and model
• models.json — provider configurations
• audit/ — audit event logs
• sessions/ — conversation history
• vault/ — encrypted secrets
• prompts/ — custom system prompts
• skills/ — local skill overrides
• index/ — semantic search index (when using az index build)

Chat with Your Project

az chat

The agent can read files, list directories, search code, and execute tools — all governed by your policy.

you> what does this project do?
  [tool: read] ok (1234 bytes)
  [tool: list] ok (450 bytes)

agentzero> This is a Rust workspace with 13 crates implementing...

Run a Security Scan

az run repo-security-audit

Scans for secrets, PII, prompt injection attempts, and unsafe patterns.

Query Your Documents

Build a semantic index of your project files, then ask questions about them in chat:

ollama pull nomic-embed-text
az index build
az chat

you> how does the policy engine work?
  [tool: query] ok (2048 bytes)

agentzero> The policy engine uses deny-by-default semantics with...

See the Document Querying guide for details.

Use as MCP Server

Add to your Claude Code or Cursor settings:

{
  "mcpServers": {
    "agentzero": {
      "command": "agentzero",
      "args": ["mcp"]
    }
  }
}

Now your editor’s AI gets policy-controlled tool access through AgentZero.

Editor Integration

Generate editor-native config files during init:

az init --private --editor vscode

This creates .vscode/tasks.json configured to launch az serve. Also works with cursor and zed.

For scripting and automation, use single-shot mode:

az chat -P "what does this project do?" --mode json

This sends one message, prints the JSON response, and exits.

Choose a Different Model

# Use a specific model
az chat --model codellama

# Use llama.cpp instead of Ollama
az chat --provider llama-cpp

# Use vLLM on a custom port
az chat --provider vllm --url http://gpu-box:8000

Search for Skills

az search "security audit"

Returns matching skills from the index with name, description, trust tier, and tags. Use --json for machine-readable output.

View Audit Summary

az audit summary

Shows a human-readable summary of audit activity: session count, total events, denied actions, and redactions applied. Use --json for structured output.

Generate a Tool

The agent can create new WASM tools during a chat session:

you> I need a tool that counts lines in all .rs files
  [APPROVE generate_tool: line-counter?] (y/yes-all/n) y
  [tool: generate_tool] ok — registered line-counter v1

agentzero> Created and registered the line-counter tool. It's now available for use.

Encryption (On by Default)

Encryption is enabled by default — you’ll be prompted for a passphrase:

az chat
# Encryption passphrase: ********
# Audit logs and sessions encrypted at rest

To disable, pass --no-encrypt.