Vanzik documentation
Install Vanzik, put it in front of the agents you already run, write the policy they answer to, and operate it in production. Coming from AgentGuard? The migration guide covers what 2.0 moves for you.
vanzik v2.0 · updated
start here
- Install and set upInstall Vanzik on Linux, macOS or Windows, start the server with a starter policy and an API key, and connect your first agent in under five minutes.
- Guard a coding agentCheck Claude Code, Cursor, Gemini CLI and Codex CLI tool calls against a policy with each agent's hooks: shell commands, file edits, web fetches and MCP tools.
- Gate MCP tool callsPut a policy in front of Claude Desktop's MCP servers in 90 seconds: every tool call is checked against YAML rules and shown live on the dashboard.
- Upgrade from AgentGuardWhat changes when you upgrade Vanzik, release by release, including moving an AgentGuard 1.x install to Vanzik 2.0 with vanzik migrate --from-agentguard.
Get started
- Setup
Install Vanzik on Linux, macOS or Windows, start the server with a starter policy and an API key, and connect your first agent in under five minutes.
- Coding agents
Check Claude Code, Cursor, Gemini CLI and Codex CLI tool calls against a policy with each agent's hooks: shell commands, file edits, web fetches and MCP tools.
- MCP gateway quickstart
Put a policy in front of Claude Desktop's MCP servers in 90 seconds: every tool call is checked against YAML rules and shown live on the dashboard.
- LLM API proxy quickstart
Route OpenAI SDK calls through Vanzik in 90 seconds, so every tool call the model emits is checked against your policy before your agent runs it.
- Migration guide
What changes when you upgrade Vanzik, release by release, including moving an AgentGuard 1.x install to Vanzik 2.0 with vanzik migrate --from-agentguard.
- FAQ
Common Vanzik questions: whether it's a sandbox, which layer to use, replicas, approvals after a restart, testing a policy, overhead and why YAML.
Integrations
- MCP gateway
vanzik-mcp-gateway sits between an MCP host (Claude Desktop, Cursor, Cline, Continue, Zed) and its servers, checking every tools/call against your policy.
- LLM API proxy
vanzik-llm-proxy forwards OpenAI and Anthropic API traffic and gates every tool call the model emits, including tool calls inside streamed responses.
- Framework adapters
Python adapters for LangChain, CrewAI, the OpenAI Agents SDK, browser-use and MCP that send every tool call through Vanzik's policy check before it runs.
- Python SDK
The vanzik Python package in depth: fail modes, timeouts, exceptions, the @guarded decorator, adapter internals and how to test code that uses it.
- Proxy architecture
How Vanzik's MCP gateway and LLM API proxy share one central server for policy, audit and approvals, and the deployment topologies they support.
Policy & approvals
- Policy reference
The complete reference for Vanzik policy YAML: scopes, allow, deny and require_approval rules, conditions, rate and cost limits, and how a check is evaluated.
- Policy configuration
Reference for the non-rule keys in a Vanzik policy file, the proxy: and notifications: settings, with each key's default, validation and runtime effect.
- Approval workflow
What happens from the moment a policy returns REQUIRE_APPROVAL until the agent proceeds or gives up: timeouts, notifications, the dashboard and retries.
- Dashboard
The web dashboard built into the vanzik binary: approving and denying from the browser, the live audit feed, session replay, login and security headers.
- Compliance reports
vanzik report turns the audit log into evidence for an auditor: what Vanzik decided over a period, whether the log was changed, and the SOC 2 criteria it bears on.
Operate
- Production deployment
What to get right before exposing Vanzik beyond localhost: an API key, TLS behind a reverse proxy, CORS, the open /v1/check endpoint and the proxies.
- Operations
Day-2 operations for Vanzik: audit log rotation, redaction, sealing and anchoring, multi-instance PostgreSQL, queue capacity, backups and restarts.
- Observability
Monitor Vanzik with the Prometheus /metrics endpoint, /api/stats and response-header timings, with example scrape and alerting rules.
- Tuning
Every setting that changes Vanzik's runtime behaviour, grouped by how you change it, from the safest to the most surgical.
- Troubleshooting
Symptom-first fixes for Vanzik: connection refused, dashboard login loops, approval links pointing at localhost, ignored policy edits, proxy errors and more.
- Service level objectives
Vanzik's service level objectives for a single-replica server: workload assumptions, latency targets, the measured baseline and how to re-run it.
Reference
- CLI
Every vanzik command and flag with its environment variable: setup, server, hook, validate, check, test, audit, audit verify, report, tenant and migrate.
- HTTP API
Every HTTP endpoint the Vanzik server exposes, with authentication requirements, request and response shapes, and status codes.
- Wire protocol
The JSON wire format of Vanzik's /v1/check endpoint: the public contract between the proxies, the SDKs and any third-party client.
- File formats
The files Vanzik writes to disk (audit log, rotated archives, replay checkpoint, vanzik.db) and the schema-version rules that keep them readable.
Security & stability
- Threat model
Who Vanzik defends against, what it trusts and what it deliberately doesn't attempt: actors, trust boundaries, fail-open versus fail-closed, and non-goals.
- Compatibility
The public surfaces Vanzik froze at 1.0 and its additive-only guarantee, which carries over to 2.x: what can change between releases and what never will.
- Deprecations
Vanzik features on their way out: when each was deprecated, when it goes, what replaces it, and the ones already removed.