Architecture

System overview

┌─────────────────────────────────────────────────┐
│                  A3T Console                     │
│          (SvelteKit SPA · console.a3t.io)        │
└──────────────────┬──────────────────────────────┘
                   │ MCP (JSON-RPC)
┌──────────────────▼──────────────────────────────┐
│                 FlowR Server                     │
│         (Rust · CNCF Workflow Runtime)           │
│                                                  │
│  ┌──────────┐ ┌──────────┐ ┌──────────────────┐ │
│  │ Workflow  │ │ Resource │ │ Trigger Manager  │ │
│  │ Engine   │ │ Registry │ │ (webhook/cron)   │ │
│  └────┬─────┘ └──────────┘ └──────────────────┘ │
│       │                                          │
│  ┌────▼─────────────────────────────────────┐   │
│  │           Task Executors (26 types)       │   │
│  │  set · call:http · call:llm · call:agent  │   │
│  │  call:lancedb · switch · for · fork · ... │   │
│  └──────────────────────────────────────────┘   │
└──────────────────┬──────────────────────────────┘
                   │
       ┌───────────┼───────────┐
       ▼           ▼           ▼
   ┌───────┐  ┌────────┐  ┌────────┐
   │ LLMs  │  │ClawDB  │  │LanceDB │
   │       │  │(Memory)│  │(Store) │
   └───────┘  └────────┘  └────────┘

Communication protocol

All communication uses MCP (Model Context Protocol) over JSON-RPC 2.0:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "id": 1,
  "params": {
    "name": "execute_flow",
    "arguments": { "name": "my-flow", "namespace": "default" }
  }
}

The FlowR server exposes MCP tools that the console, CLI, and other MCP clients can call.

Execution model

1. Workflow registered — YAML parsed and stored
2. Execution requested — instance created, queued
3. Steps execute sequentially — each step runs its task executor
4. Async completion — poll get_execution_status for result
5. Output returned — final step output is the workflow result

Workflows are fire-and-forget with polling. The execute_flow call returns immediately with an instance ID. Use get_execution_status to check progress.

Resource model

Resources are named connections to external systems. They’re referenced by steps:

- name: query-db
  type: call:lancedb
  resource: a3t/a3t-platform-db    # namespace/name reference
  config:
    operation: query
    table: users

Resource types: openai, anthropic, ollama, postgres, mysql, lancedb, redis, qdrant, mcp, http, clawdb, agent.

Namespace isolation

All entities (workflows, resources, triggers) are scoped to a namespace. The default namespace is default. A3T platform entities use the a3t namespace.