copilotkit--copilotkit/dev-docs/architecture/setup-runtime.md

Runtime / Backend Setup Guide

This guide shows how to set up the CopilotKit backend — from minimal to fully configured with all optional extension points.

---

What Talks to What

graph TB
    subgraph Frontends
        React["React App"]
        Angular["Angular App"]
        Vanilla["Vanilla JS"]
    end

    subgraph Your Server
        EP["Express / Hono<br/><i>Endpoint handler</i>"]
        BM["beforeRequestMiddleware<br/><i>(optional)</i>"]
        RT["<b>CopilotRuntime</b>"]
        AM["afterRequestMiddleware<br/><i>(optional)</i>"]
        Runner["<b>AgentRunner</b><br/><i>InMemory (default)<br/>or SQLite</i>"]
        TS["TranscriptionService<br/><i>(optional)</i>"]
    end

    subgraph Agents
        A1["Agent 1<br/><i>LangGraph</i>"]
        A2["Agent 2<br/><i>CrewAI</i>"]
        A3["Agent 3<br/><i>Custom</i>"]
    end

    React -->|HTTP| EP
    Angular -->|HTTP| EP
    Vanilla -->|HTTP| EP
    EP --> BM
    BM --> RT
    RT --> AM
    RT --> Runner
    RT --> TS
    Runner -->|AG-UI events| A1
    Runner -->|AG-UI events| A2
    Runner -->|AG-UI events| A3

---

Minimal Setup (Express)

1. Install

npm install @copilotkit/runtime express

2. Create the runtime

// server.ts
import express from "express";
import { CopilotRuntime } from "@copilotkit/runtime";
import { createCopilotEndpointExpress } from "@copilotkit/runtime/express";

const app = express();

const runtime = new CopilotRuntime({
  agents: {
    default: myAgent, // Any AbstractAgent implementation
  },
});

app.use("/api/copilotkit", createCopilotEndpointExpress({ runtime }));

app.listen(3000);

That's it. The endpoint handler creates these routes automatically:

Route	Method	What it does
/api/copilotkit/info	GET	Returns list of available agents
/api/copilotkit/agent/:agentId/run	POST	Run an agent (returns SSE stream)
/api/copilotkit/agent/:agentId/connect	POST	Connect/reconnect to a thread
/api/copilotkit/agent/:agentId/stop/:threadId	POST	Stop a running agent
/api/copilotkit/transcribe	POST	Audio transcription (if configured)

sequenceDiagram
    participant Client as Frontend
    participant EP as Express Router
    participant RT as CopilotRuntime
    participant Agent as AI Agent

    Client->>EP: GET /api/copilotkit/info
    EP->>RT: List agents
    RT-->>Client: [{ id: "default", description: "..." }]

    Client->>EP: POST /agent/default/run
    EP->>RT: handleRunAgent({ agentId: "default" })
    RT->>Agent: runner.run()
    Agent-->>Client: SSE: TEXT_MESSAGE_START
    Agent-->>Client: SSE: TEXT_MESSAGE_CONTENT
    Agent-->>Client: SSE: TEXT_MESSAGE_END
    Agent-->>Client: SSE: RUN_FINISHED

---

Hono Setup

import { Hono } from "hono";
import { CopilotRuntime } from "@copilotkit/runtime";
import { createCopilotEndpointHono } from "@copilotkit/runtime/hono";

const app = new Hono();

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
});

app.route("/api/copilotkit", createCopilotEndpointHono({ runtime }));

export default app;

---

Using the Built-in Agent

CopilotKit includes a built-in agent powered by the Vercel AI SDK:

import { CopilotRuntime } from "@copilotkit/runtime";
import { BuiltInAgent } from "@copilotkit/runtime/v2";

const agent = new BuiltInAgent({
  model: "openai/gpt-4o",
  systemPrompt: "You are a helpful shopping assistant.",
});

const runtime = new CopilotRuntime({
  agents: { default: agent },
});

---

Lazy-Loaded Agents

Agents can be a Promise — useful for dynamic loading:

const runtime = new CopilotRuntime({
  agents: loadAgents(), // Returns Promise<Record<string, AbstractAgent>>
});

async function loadAgents() {
  const config = await fetchConfig();
  return {
    default: new BuiltInAgent({ model: config.model }),
    research: new HttpAgent({ url: config.researchAgentUrl }),
  };
}

---

All CopilotRuntime Options

const runtime = new CopilotRuntime({
  // Required: map of agent IDs to agent instances
  agents: {
    default: defaultAgent,
    research: researchAgent,
    coding: codingAgent,
  },

  // Optional: how agents are executed
  runner: new InMemoryAgentRunner(), // default

  // Optional: audio → text
  transcriptionService: myTranscriptionService,

  // Optional: intercept requests before processing
  beforeRequestMiddleware: async ({ request, path }) => {
    console.log(`[${path}] Request received`);
    // Return modified request, or void to pass through
  },

  // Optional: run after response is prepared
  afterRequestMiddleware: async ({ response, path }) => {
    console.log(`[${path}] Response sent`);
  },
});

graph TB
    subgraph "CopilotRuntime Options"
        direction TB

        subgraph "Required"
            AGENTS["agents<br/><i>Record&lt;string, AbstractAgent&gt;</i>"]
        end

        subgraph "Optional"
            RUNNER["runner<br/><i>AgentRunner</i><br/><i>Default: InMemoryAgentRunner</i>"]
            TS["transcriptionService<br/><i>TranscriptionService</i>"]
            BM["beforeRequestMiddleware<br/><i>(request, path) → Request | void</i>"]
            AM["afterRequestMiddleware<br/><i>(response, path) → void</i>"]
        end
    end

---

AgentRunner: How Agents Execute

The AgentRunner is the abstraction that actually executes agents. It manages threads, streaming, and agent lifecycle.

graph TB
    subgraph "AgentRunner (Abstract)"
        RUN["run(request)<br/><i>Execute agent, return Observable&lt;BaseEvent&gt;</i>"]
        CONNECT["connect(request)<br/><i>Reconnect to existing thread</i>"]
        RUNNING["isRunning(request)<br/><i>Check if thread is active</i>"]
        STOP["stop(request)<br/><i>Abort a running thread</i>"]
    end

    subgraph Implementations
        IM["<b>InMemoryAgentRunner</b><br/><i>Default — in-process, ephemeral</i>"]
        SQ["<b>SQLiteAgentRunner</b><br/><i>Persistent state on disk</i>"]
        CU["<b>Your Custom Runner</b><br/><i>Extend AgentRunner</i>"]
    end

    IM --> RUN
    SQ --> RUN
    CU --> RUN

InMemoryAgentRunner (default)

Stores agent threads in memory. Simple, no persistence. Threads are lost on server restart.

import { InMemoryAgentRunner } from "@copilotkit/runtime";

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
  runner: new InMemoryAgentRunner(), // This is the default
});

SQLiteAgentRunner (persistent)

Stores agent threads in SQLite. Survives restarts.

import { SQLiteAgentRunner } from "@copilotkit/sqlite-runner";

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
  runner: new SQLiteAgentRunner({ dbPath: "./agent-state.db" }),
});

Custom Runner

import { AgentRunner } from "@copilotkit/runtime";
import { Observable } from "rxjs";

class RedisAgentRunner extends AgentRunner {
  async run(request) {
    // Store in Redis, return event stream
    return new Observable((subscriber) => {
      // ... your implementation
    });
  }

  async connect(request) {
    // Reconnect to existing Redis-stored thread
  }

  async isRunning(request) {
    // Check Redis for active thread
  }

  async stop(request) {
    // Signal thread to stop
  }
}

---

Middleware

Middleware lets you intercept requests before and after processing.

Before Request Middleware

Runs before any handler. Use it for auth, logging, request modification.

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
  beforeRequestMiddleware: async ({ request, path, runtime }) => {
    // Example: verify auth token
    const token = request.headers.get("authorization");
    if (!token) {
      return new Response("Unauthorized", { status: 401 });
    }

    // Example: add user context to request
    const user = await verifyToken(token);
    request.headers.set("x-user-id", user.id);

    // Return modified request (or void to pass through)
    return request;
  },
});

After Request Middleware

Runs after the response is prepared but before it's sent.

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
  afterRequestMiddleware: async ({ response, path, runtime }) => {
    // Example: log responses
    console.log(`[${path}] Response status: ${response.status}`);

    // Example: add custom headers
    // (Note: response may be SSE stream)
  },
});

sequenceDiagram
    participant Client
    participant Before as beforeRequestMiddleware
    participant Handler as Route Handler
    participant After as afterRequestMiddleware

    Client->>Before: HTTP Request
    alt Middleware rejects
        Before-->>Client: 401 Unauthorized
    else Middleware passes
        Before->>Handler: (modified) Request
        Handler->>After: Response
        After-->>Client: Final Response
    end

---

Transcription Service

Enable audio-to-text transcription:

import { TranscriptionService } from "@copilotkit/runtime";

class OpenAITranscription extends TranscriptionService {
  async transcribeFile({ audioFile, mimeType, size }) {
    const formData = new FormData();
    formData.append("file", audioFile);
    formData.append("model", "whisper-1");

    const response = await fetch(
      "https://api.openai.com/v1/audio/transcriptions",
      {
        method: "POST",
        headers: { Authorization: `Bearer ${process.env.OPENAI_API_KEY}` },
        body: formData,
      },
    );

    const result = await response.json();
    return result.text;
  }
}

const runtime = new CopilotRuntime({
  agents: { default: myAgent },
  transcriptionService: new OpenAITranscription(),
});

The /transcribe endpoint is only active when transcriptionService is configured.

---

Request Flow Detail

sequenceDiagram
    participant Client as Frontend
    participant CORS as CORS Handler
    participant Before as Before Middleware
    participant Router as Route Handler
    participant Runtime as CopilotRuntime
    participant Runner as AgentRunner
    participant Agent as AI Agent
    participant After as After Middleware

    Client->>CORS: POST /agent/default/run
    CORS->>Before: Check middleware
    Before->>Router: Forward request
    Router->>Runtime: handleRunAgent()

    Note over Runtime: 1. Resolve agents (await if Promise)
    Note over Runtime: 2. Find agent by ID
    Note over Runtime: 3. Clone agent (avoid shared state)
    Note over Runtime: 4. Parse RunAgentInput from body
    Note over Runtime: 5. Set messages, state, threadId

    Runtime->>Runner: runner.run({ agent, input })
    Runner->>Agent: agent.runAgent(input)

    loop SSE Stream
        Agent-->>Client: event: TEXT_MESSAGE_CONTENT
    end

    Agent-->>Client: event: RUN_FINISHED
    Router->>After: Response complete

---

Full Example: Express + Multiple Agents + Middleware

import express from "express";
import { CopilotRuntime } from "@copilotkit/runtime";
import { createCopilotEndpointExpress } from "@copilotkit/runtime/express";
import { BuiltInAgent } from "@copilotkit/runtime/v2";
import { SQLiteAgentRunner } from "@copilotkit/sqlite-runner";

const app = express();

// Create agents
const generalAgent = new BuiltInAgent({
  model: "openai/gpt-4o",
  systemPrompt: "You are a helpful assistant.",
});

const codeAgent = new BuiltInAgent({
  model: "openai/gpt-4o",
  systemPrompt: "You are a coding expert. Always provide code examples.",
});

// Create runtime with all optional features
const runtime = new CopilotRuntime({
  agents: {
    default: generalAgent,
    coding: codeAgent,
  },

  // Persistent agent state
  runner: new SQLiteAgentRunner({ dbPath: "./data/agents.db" }),

  // Auth middleware
  beforeRequestMiddleware: async ({ request, path }) => {
    const token = request.headers.get("authorization")?.replace("Bearer ", "");
    if (!token) {
      return new Response(JSON.stringify({ error: "Unauthorized" }), {
        status: 401,
        headers: { "Content-Type": "application/json" },
      });
    }
    // Validate token...
  },

  // Logging middleware
  afterRequestMiddleware: async ({ path }) => {
    console.log(`[CopilotKit] ${new Date().toISOString()} ${path}`);
  },
});

// Mount CopilotKit endpoints
app.use("/api/copilotkit", createCopilotEndpointExpress({ runtime }));

app.listen(3000, () => {
  console.log("Server running on :3000");
  console.log("CopilotKit endpoints at /api/copilotkit/*");
});