cloudflare/agents is a comprehensive, officially maintained agent SDK from Cloudflare; MCP server/client support is one feature among many rather than the repo's sole purpose. Documentation and examples are extensive, but the project does not accept external PRs and has no standalone MCP install configuration separate from writing code.
cloudflare/agents is Cloudflare's officially maintained TypeScript SDK for building persistent, stateful AI agents on top of Durable Objects — each agent gets its own state, storage, and lifecycle, hibernating when idle and waking on demand. It is not itself a single MCP server with a fixed tool list; rather it's an application framework that lets developers make their own Agent act as an MCP server (exposing it to MCP clients) or as an MCP client (connecting to other MCP servers), over HTTP, SSE, RPC, and with elicitation support. Beyond MCP, the SDK bundles synced persistent state, type-safe callable methods (RPC via @callable()), sub-agent composition, scheduling, WebSockets, AI chat with resumable streaming, WebMCP, durable workflows with human-in-the-loop approval, email send/receive, a voice pipeline, browser-side agents, Code Mode (LLMs write executable code instead of one tool call at a time), sandboxed code execution, and x402 pay-per-call support. The repo is a pnpm/Nx monorepo containing the core agents package plus companion packages (ai-chat, think, codemode, shell, voice, worker-bundler, hono-agents) and 30+ standalone examples, several of them specifically demonstrating MCP usage (mcp, mcp-client, mcp-server, mcp-worker, webmcp). The project does not accept external pull requests but welcomes bug reports and discussion via GitHub Issues/Discussions.
Run npm install agents in an existing Cloudflare Workers project, or scaffold a new one with npm create cloudflare@latest -- --template cloudflare/agents-starter. Declare your Agent class as a Durable Object and configure the matching durable_objects binding plus a SQLite migration (new_sqlite_classes) in wrangler.jsonc, along with an appropriate compatibility_date and the nodejs_compat flag. To expose an Agent as an MCP server or connect it as an MCP client, follow the patterns in the repo's examples/mcp* directories — there is no separate one-step MCP installer or client config file provided.