Dispatch

@rubrion/mastra-cloudflare 0.2: Mastra agents on Workers, from npm

What began as a starter in June is now a library: an OpenAI-compatible agent server, D1 memory and model helpers, in one npm package.

Most agent stacks assume a server running all day, and every project ends up rewriting the same glue: an API, routing between agents, memory. @rubrion/mastra-cloudflare is that glue, for Mastra agents on Cloudflare Workers.

From starter to library

  • June 13: the repository started as a starter: a Mastra agent on Workers with D1, to copy and edit.
  • October 20: every project we built on it copied the same files, so we turned it into a library: 0.1.0, published on npm. The rubry template and rubry.ai run on it.
  • October 29: 0.2.0, with per-environment limits and token usage in every reply.
  • October 30: 0.2.1, extra headers on remote models, so a platform can pass who's asking to an agent that runs elsewhere. That's how Equânime keeps each person's analyses apart on rubry.ai.

What's in the package

  • createAgentServer: an OpenAI-compatible server (/v1/models, /v1/chat/completions with real SSE streaming) as a Worker fetch handler, with no HTTP framework. model: "auto" routes each message to the right agent, or splits it across several.
  • defineAgents: agents as plain config, built lazily once per isolate, because bindings and secrets can only be read inside a request.
  • d1Memory: Mastra Memory on D1, with no vector store to pay for.
  • workersAI, openAICompatible, providerModel: models from Workers AI, any OpenAI-compatible endpoint, or the provider each request carries.
  • Re-exports Agent, createTool, z and RequestContext, so a project needs only this package.

A Worker in 20 lines

npm install @rubrion/mastra-cloudflare
import { env } from 'cloudflare:workers';
import { createAgentServer, createTool, d1Memory, defineAgents, z } from '@rubrion/mastra-cloudflare';

const time = createTool({
	id: 'current_time',
	description: 'The current date and time.',
	inputSchema: z.object({}),
	execute: async () => ({ now: new Date().toISOString() }),
});

const agents = defineAgents(
	[
		{ id: 'writer', name: 'Writer', description: 'Writes and edits text.', instructions: 'Write what is asked, briefly.' },
		{ id: 'helper', name: 'Helper', description: 'General questions; knows the time.', instructions: 'Be useful.', tools: { current_time: time } },
	],
	{ memory: d1Memory(() => env.MEMORY) },
);

export default createAgentServer({
	agents,
	apiKey: (e: Env) => e.API_KEY,
	provider: (e: Env) => ({ baseURL: 'https://api.openai.com/v1', apiKey: e.OPENAI_API_KEY, model: 'gpt-5-mini' }),
});

New in 0.2

  • maxAgents (how many agents auto may split a request across, default 3) and maxSteps (tool rounds per reply, default 5) take a number or (env) => number, so a public demo can be stricter than production without a second build.
  • Token usage in every reply: OpenAI's usage field, summed over the planner, every agent and every tool round, in the JSON body or on the final streamed chunk. rubry.ai's daily budget is computed from it.

Links

Building agents on Cloudflare and want a hand? Let's talk.

By post

Subscribe to the newsletter

Delivered to your inbox. Unsubscribe anytime.