All docs

Build an agent

Stateful agents, chat assistants and MCP servers on Durable Objects, with env.AI for models, Vector memory, approvals and connectors — from a Studio starter or your own code.

Open in the console

An agent on RandallFlare is an app: JavaScript or TypeScript on the edge, with randallflare:agents for the stateful parts (one Durable Object per agent instance), env.AI for models, a Vector index for memory, Workflows for long jobs and Approvals when a person must decide. There is no separate agent product to pay for; agents run and are billed like any other app.

Start from a template

Starter	What it contains	Bindings it expects
Agent (chat + schedule)	an `AIChatAgent` that streams replies over WebSocket behind a small page, plus a scheduled `Agent` that keeps state and reports it	`AI` (ai); `ASSISTANT`, (Durable Objects)

MONITOR

// Chat (AIChatAgent): one conversation per name
const ws = new WebSocket("wss://assistant.edge.randallflare.com/agents/assistant/alice");
ws.onopen = () => ws.send(JSON.stringify({ type: "rf_chat_history" }));          // → { type: "rf_chat_messages", messages }
ws.send(JSON.stringify({ type: "rf_chat_send", content: "Summarise today" }));   // → rf_chat_chunk { delta }… then rf_chat_done { message }

// Without WebSockets: POST …/chat for a text/event-stream reply, GET …/messages for the history
await fetch("https://assistant.edge.randallflare.com/agents/assistant/alice/chat", { method: "POST", body: JSON.stringify({ content: "Hi" }) });

// State (Agent): every change is pushed as rf_agent_state
const mon = new WebSocket("wss://assistant.edge.randallflare.com/agents/monitor/prod");
mon.onmessage = (e) => { const m = JSON.parse(e.data); if (m.type === "rf_agent_state") render(m.state); };
mon.send(JSON.stringify({ type: "rf_agent_call", id: 1, method: "tick", args: [] }));   // methods listed in static callable

Build an agent

Start from a template

The pieces

Talk to it

Keep it safe