CopilotKit Runtime

@copilotkit/runtime is the server half of CopilotKit: it accepts AG-UI protocol requests, dispatches them to an AbstractAgent (built-in or external), runs the stream through an AgentRunner, and responds as Server-Sent Events.

This SKILL.md is the index. Read the reference under references/ that matches your task — do not try to absorb the whole package from this file.

Mental Model — the three dictionaries you hand to `CopilotRuntime`

new CopilotRuntime({
  agents, // Record<string, AbstractAgent>     — see wiring-external-agents or built-in-agent
  runner, // AgentRunner (optional)            — see agent-runners
  intelligence, // CopilotKitIntelligence (optional) — see intelligence-mode (auto-wires runner)
  mcpApps, // McpAppsConfig (optional)          — see wiring-mcp-apps-middleware
  a2ui, // A2UIConfig (optional)             — see packages/a2ui-renderer skill
  hooks, // { onRequest, onBeforeHandler }    — see middleware
  beforeRequestMiddleware,
  afterRequestMiddleware, // legacy — see middleware
  transcription, // TranscriptionService (optional)  — see transcription
});

You then mount it:

import { createCopilotRuntimeHandler } from "@copilotkit/runtime/v2";
const handler = createCopilotRuntimeHandler({
  runtime,
  basePath: "/api/copilotkit",
});
export default { fetch: handler };

When to load which reference

Task	Reference
Mounting on any fetch-native server (Cloudflare Workers, Bun, Deno, Vercel Edge, Next.js App Router, React Router v7, TanStack Start) or delegating from Express/Node	`references/setup-endpoint.md`
Auth / logging / rate-limit / request-scoped guards via `hooks.onRequest` / `hooks.onBeforeHandler` (preferred) or legacy `beforeRequestMiddleware` / `afterRequestMiddleware`	`references/middleware.md`
Choosing between `InMemoryAgentRunner`, `SqliteAgentRunner`, or a custom subclass — including thread-locking semantics and the runner/Intelligence mutual exclusion	`references/agent-runners.md` (+ `-in-memory.md`, `-sqlite.md`, `-custom.md` for backend-specific detail)
Enabling durable threads + realtime websocket via CopilotKit Cloud (Intelligence is Cloud-only, not self-hostable)	`references/intelligence-mode.md`
Voice transcription — implementing a `TranscriptionService` subclass for the `/transcribe` endpoint	`references/transcription.md`
Instantiating `BuiltInAgent` — Simple Mode (classic) or Factory Mode with TanStack AI (preferred AG-UI-compliant default), AI SDK, or custom factory	`references/built-in-agent.md` (+ `-factory-modes.md`, `-helper-utilities.md`, `-model-identifiers.md`)
Defining server-side tools via `defineTool` for `BuiltInAgent.config.tools` (Simple Mode only)	`references/server-side-tools.md`
Wiring an external agent framework into `CopilotRuntime({ agents })`	`references/wiring-external-agents.md` (index) + per-framework refs (`wiring-mastra.md`, `wiring-langgraph.md`, `wiring-crewai-crews.md`, `wiring-crewai-flows.md`, `wiring-pydantic-ai.md`, `wiring-adk.md`, `wiring-llamaindex.md`, `wiring-agno.md`, `wiring-aws-strands.md`, `wiring-ms-agent-framework.md`, `wiring-ag2.md`, `wiring-a2a.md`)
Wiring MCP Apps (runtime-level middleware, not an agent)	`references/wiring-mcp-apps-middleware.md`

Invariants and gotchas (load-once, before any reference)

createCopilotRuntimeHandler is the canonical primitive. createCopilotExpressHandler / createCopilotHonoHandler exist but are avoid at all costs — delegate from Express/Hono routes to the fetch primitive instead.
publicLicenseKey is the canonical provider-side field. publicApiKey is a deprecated alias — expect to see it in legacy code, emit the canonical name in new code.
Intelligence mode auto-wires IntelligenceAgentRunner. Passing both runner and intelligence to CopilotRuntime is rejected at construction.
Intelligence mode targets CopilotKit Cloud (api.cloud.copilotkit.ai) and is not self-hostable.
hooks.onRequest runs before beforeRequestMiddleware (hook-based middleware wins for Response short-circuits). beforeRequestMiddleware runs after hooks.onRequest (see fetch-handler.ts:136-147).
identifyUser (Intelligence) does not forward thrown Response objects — convert to 500. Gate auth rejection in hooks.onRequest, which does forward Responses.
agents__unsafe_dev_only and selfManagedAgents are dev-only aliases of each other; do not reach for them in production. Either signals that the SPA is in dev mode.

Reading order for a first-time reader

setup-endpoint — the primitive.
built-in-agent or pick one from wiring-external-agents — the agent.
agent-runners — production persistence choice.
Optional: middleware, intelligence-mode, server-side-tools, transcription.