Introduction

Vystak is to AI agents what Pulumi is to cloud infrastructure. Define your agent once in YAML or Python, and Vystak generates the framework code, provisions the infrastructure, and deploys the agent — to Docker, Azure Container Apps, or any future platform.

Vystak builds nothing. It wires everything.

What you can do with Vystak

• Define agents declaratively — one YAML or Python file describes the model, tools, sessions, channels, and where to run.
• Deploy anywhere — Docker locally, Azure Container Apps in production. Same definition, different target.
• OpenAI-compatible API — every agent exposes /v1/chat/completions and /v1/responses out of the box. Drop-in replacement for any OpenAI client.
• Multi-agent collaboration — built-in A2A protocol, gateway routing, and registry. Agents discover and call each other natively.
• Persistence built in — Postgres sessions, long-term memory, all auto-provisioned alongside the agent.
• Hash-based change detection — vystak apply only redeploys what changed.

How it works

   ┌──────────────────────┐
   │  vystak.yaml or .py  │   ← Define once
   └──────────┬───────────┘
              │
   ┌──────────▼───────────┐
   │   Framework adapter  │   ← Generates LangGraph + FastAPI code
   │       (LangChain)    │
   └──────────┬───────────┘
              │
   ┌──────────▼───────────┐
   │  Platform provider   │   ← Provisions infra, builds image, deploys
   │  (Docker, Azure,...) │
   └──────────┬───────────┘
              │
   ┌──────────▼───────────┐
   │   Running agent      │   ← /invoke /stream /v1/chat/completions
   └──────────────────────┘

Three independent choices for every deployment:

• Framework adapter — how the agent thinks (LangChain/LangGraph today, others coming)
• Platform provider — where it runs (Docker, Azure Container Apps)
• Channel adapter — how users reach it (REST API, Slack, webhook)

Any combination works. The agent definition doesn't change — only the platform target does.

Core concepts

Concept	What it is
Agent	The deployable unit — model, tools, sessions, channels
Model	Which LLM and how to call it (Anthropic, OpenAI-compatible, MiniMax)
Provider	A cloud account or service (docker, azure, anthropic)
Platform	Where the agent runs (docker, container-apps)
Service	Backing infrastructure (Postgres, Redis, Qdrant)
Channel	How users reach the agent (REST, Slack, webhook)

Each gets its own page — for now, the Agents page covers the basics. The other concept pages have placeholders we'll expand soon.

What's next

• Installation — install the CLI and Python packages
• Quickstart — deploy your first agent in five minutes
• Agents — the agent schema in depth