agents-sdk

Extend AIChatAgent from the agents package, install @cloudflare/ai-chat and the AI SDK, and implement onChatMessage with streamText. Configure a Durable Object binding and a new_sqlite_classes migration in wrangler.jsonc, then connect with useAgentChat on the client.

Use this.schedule with a delay, date, or cron expression for one-time and calendar-based work, or this.scheduleEvery for fixed intervals. The agent invokes the named callback method when the schedule fires, with retries configurable via this.retry.

Each agent class needs a Durable Object binding, a new_sqlite_classes migration entry, and the nodejs_compat compatibility flag. Never edit old migrations; add a new tag for new classes, and do not enable experimentalDecorators in tsconfig since it breaks @callable.

Yes, the SDK includes an MCP client API for connecting to remote servers with OAuth support, and createMcpHandler for building your own MCP servers. It supports Streamable HTTP, SSE, and RPC transports, with guidance for securing servers.

In-flight work is lost on eviction unless you use runFiber, which checkpoints progress to SQLite via ctx.stash. On restart, onFiberRecovered receives the checkpoint so you can resume; for multi-step pipelines with automatic retries, use Workflows instead.

Think is an experimental higher-level class that handles the streamText loop, tool execution, and persistence automatically via getModel and getSystemPrompt overrides. AIChatAgent gives you full control inside onChatMessage but requires wiring the streaming loop yourself.