Compare

Copilotz vs the Vercel AI SDK

The AI SDK is a toolkit for calling models and building chat interfaces. Copilotz is a runtime for the whole app. They work at different layers; here's what that means in practice, with a source for every claim.

Checked on 12 November 2026.

In short

The Vercel AI SDK is a TypeScript toolkit for AI apps and agents: AI SDK Core calls models, with tools, structured output and agents, and AI SDK UI gives you streaming chat hooks for React, Vue, Svelte and Angular. It's the most widely used library of its kind, now on version 7, with new releases almost every day.

Copilotz works one layer up. It's a runtime for the whole app: its data, the Actions that change it, who may call them, how it shows on a page or in a model's prompt, and a record of every write. The AI SDK leaves those to you, by design: storage, users, permissions and serving MCP are your code.

So the question is how much you want to own. Pick the AI SDK to add model calls and a chat UI to an app you already have. Pick Copilotz when you want the app, its rules and its agents to come from one place.

Choose the AI SDK if…

You want to add model calls, tools or a chat UI to an app you already have, and keep your own database, auth and hosting.
You want the widest choice of models behind one API, with Vercel's AI Gateway as the default route.
You want official streaming chat hooks for React, Vue, Svelte and Angular.
You want the most adopted library of its kind: about 44.8 million weekly npm downloads, and a stable 7.x release.
You need an MCP client, to use the tools of outside MCP servers.

Choose Copilotz if…

You'd rather not build what the SDK leaves to you: storage, users and permissions, a record of who did what, and an MCP server.
Your agents and outside assistants should act through the same Actions as your web page, under the same rules.
You want your app served over MCP, with widgets in ChatGPT and Claude, and on WhatsApp, from the same plugins.
You want offline tests of the whole app, the agent's turns included, with a scripted model, and a release candidate is fine for you.

Side by side

Each claim links to its source. Numbers move: these were read on 12 November 2026.

Copilotz vs the Vercel AI SDK
Dimension	Copilotz	AI SDK
What it is	Copilotz:A runtime for apps that people and AI use alike. You write the app once, as plugins: its data, its Actions, who may call them, and how it shows.^[1]	AI SDK:A TypeScript toolkit for AI apps and agents: AI SDK Core calls models, with tools and agents; AI SDK UI builds chat interfaces.^[2]
Runs on	Copilotz:Node 24 or later, Deno and Bun. Built on web standards, for Cloudflare Workers and the browser too.^[1]^[3]	AI SDK:Any JavaScript environment, such as Node.js, Deno and the browser. Version 7 needs Node 22 or later and is ESM-only.^[4]^[5]
Agents, tools and workflows	Copilotz:Agents are settings: instructions, a model and tools. A tool is a thin caller of an Action, the same one the web page and MCP call. Processors react to events; schedules and routines run work on a timer.^[1]^[3]	AI SDK:generateText and streamText with tools, multi-step with stopWhen, and the ToolLoopAgent class. Durable workflows come from a separate package, the Workflow DevKit.^[6]^[7]^[8]
Memory	Copilotz:The history plugin summarizes long chats; the memory plugin remembers across them, by meaning with an embeddings model, or by words without one.^[3]^[9]	AI SDK:Not kept by the SDK: you store messages yourself, as its persistence guide shows. Long-term memory comes through providers such as Mem0 or Letta, or a tool of your own.^[10]^[11]
MCP (client and server)	Copilotz:Server: the mcp plugin offers the app's tools to ChatGPT, Claude and any MCP client, and draws its web Views inside them as widgets. No MCP sign-in yet: each conversation comes in as a contact. Client: not offered; the docs don't use outside MCP servers as tools.^[3]^[12]^[13]	AI SDK:Client: createMCPClient (@ai-sdk/mcp), over HTTP, SSE or stdio. Server: not in the SDK; Vercel's separate mcp-handler package hosts one.^[14]^[15]
Auth and access rules	Copilotz:Built in, under MIT. Every call names who acts, for whom and where. Each Action says who may call it (allow), asked of an agent and of whom it works for. Rules at the data filter every read and check every write, whatever the way in.^[1]^[3]	AI SDK:No built-in users, roles or data rules: those are your code. Tools can ask for approval before they run, and an optional package moves approval rules into Open Policy Agent.^[6]^[10]^[16]
Persistence	Copilotz:One store for everything: in memory, SQLite or PostgreSQL, held to one conformance suite. Moving between them changes one line.^[17]	AI SDK:None built in: you bring the database.^[10]
Observability	Copilotz:Every write and Action call is an event naming who acted, for whom and where. Metrics follow OpenTelemetry's model, exported over OTLP or as Prometheus text.^[3]^[18]	AI SDK:OpenTelemetry, through the @ai-sdk/otel package.^[19]
Testing	Copilotz:The whole app, tested offline with a scripted model and a manual clock, and no API key. copilotz/testing is one test API for Node, Deno and Bun.^[1]^[3]	AI SDK:Mock models and helpers in ai/test (MockLanguageModelV4, for one) and simulateReadableStream, to test without calling a provider.^[20]
Deployment	Copilotz:You host it: a Node web host and a store from the environment, SQLite or PostgreSQL, with a lease so a deploy hands over without two writers.^[21]	AI SDK:A library: it runs inside your app, on Vercel or any other host.^[5]
Channels	Copilotz:A web page, a terminal, a REST API and client, MCP with widgets, WhatsApp (Meta's Cloud API), voice calls (Twilio) and a chat on your site.^[3]	AI SDK:None in the SDK. Vercel's separate Chat SDK builds bots for Slack, Teams, Google Chat, Discord, Telegram and more.^[22]
License	Copilotz:MIT.^[3]	AI SDK:Apache 2.0.^[5]
Maturity	Copilotz:A release candidate: 1.0.0-rc.45, on 12 November 2026. 1.0 isn't out yet. A young project with a small community.^[23]	AI SDK:Stable: version 7 since 25 June 2026; ai 7.0.171 on 12 November 2026. About 27,900 GitHub stars and 44.8 million npm downloads a week.^[24]^[25]^[26]^[27]
Ecosystem	Copilotz:Model providers through four protocols (OpenAI, OpenAI-compatible, Anthropic and Gemini). Its own plugins for sign-in, spaces, memory, skills, routines, boards, documents and billing. Few outside integrations so far.^[3]^[28]	AI SDK:Dozens of official and community model providers, the AI Gateway as the default route, and chat hooks for React, Vue, Svelte and Angular.^[26]^[29]^[30]

A tool, in each

The AI SDK passes tools to a model call and runs the loop. Copilotz declares the tool in a plugin, and the tool calls an Action that the web page and MCP also call, under the same rules.

AI SDK: a tool in a model call

import { z } from 'zod';import { generateText, tool, isStepCount } from 'ai'; const result = await generateText({  model: "anthropic/claude-sonnet-5.5",  tools: {    weather: tool({      description: 'Get the weather in a location',      inputSchema: z.object({        location: z.string().describe('The location to get the weather for'),      }),      execute: async ({ location }) => ({        location,        temperature: 72 + Math.floor(Math.random() * 21) - 10,      }),    }),  },  stopWhen: isStepCount(5),  prompt: 'What is the weather in San Francisco?',});

From ai-sdk.dev: Tool calling

Copilotz: salon.ts (excerpt)

// From the README's salon.ts: one Action, three callers (excerpt).actions: {  "booking.make": defineAction({    input: booking,    // Asked of every caller, tools included: people, agents working for them, and contacts.    allow: peopleOrContacts,    run: (ctx, input) => ctx.db.collection("booking").insertOne(input),  }),},views: {  // What a model reads after it books, the app's own or one over MCP.  "tool:book": (_, ctx) => ctx.show("bookings", {}),},resources: {  // The rules: a booking lives in a space, and belongs to the chat it was made in.  shared: { booking: { part: "thread" } },  tools: {    book: tool({      description: "Books a slot, like 'Wed 15:00'.",      input: { type: "object", properties: { slot: { type: "string" } }, required: ["slot"] },      run: ({ slot }, ctx) => ctx.call("booking.make", { slot, in: ctx.in, thread: ctx.thread }),    }),  },  agents: {    concierge: { instructions: "Book slots.", model: "openai/gpt-5.4-mini", tools: ["book"] },  },  // An MCP server at /mcp/salon. No sign-in: each conversation is a contact, in its own chat.  mcp: { salon: { in: "desk", tools: { book: {} } } },},

From GitHub: copilotzhq/copilotz (README)

Sources

What we say about AI SDK comes from its own docs, repository, npm and announcements; what we say about Copilotz, from its docs and repository. Something wrong or out of date? Write to hello@copilotz.com and we'll fix it.

See it for yourself.

The guide builds one app, for people and for AI, in ten chapters, and its tests run offline.

Read the guide

In short

Choose the AI SDK if…

You want to add model calls, tools or a chat UI to an app you already have, and keep your own database, auth and hosting.

You want the widest choice of models behind one API, with Vercel's AI Gateway as the default route.

You want official streaming chat hooks for React, Vue, Svelte and Angular.

You want the most adopted library of its kind: about 44.8 million weekly npm downloads, and a stable 7.x release.

You need an MCP client, to use the tools of outside MCP servers.

Choose Copilotz if…

You'd rather not build what the SDK leaves to you: storage, users and permissions, a record of who did what, and an MCP server.

Your agents and outside assistants should act through the same Actions as your web page, under the same rules.

You want your app served over MCP, with widgets in ChatGPT and Claude, and on WhatsApp, from the same plugins.

You want offline tests of the whole app, the agent's turns included, with a scripted model, and a release candidate is fine for you.

What it is

Copilotz:A runtime for apps that people and AI use alike. You write the app once, as plugins: its data, its Actions, who may call them, and how it shows.^[1]

AI SDK:A TypeScript toolkit for AI apps and agents: AI SDK Core calls models, with tools and agents; AI SDK UI builds chat interfaces.^[2]

Runs on

Copilotz:Node 24 or later, Deno and Bun. Built on web standards, for Cloudflare Workers and the browser too.^[1]^[3]

AI SDK:Any JavaScript environment, such as Node.js, Deno and the browser. Version 7 needs Node 22 or later and is ESM-only.^[4]^[5]

Agents, tools and workflows

Copilotz:Agents are settings: instructions, a model and tools. A tool is a thin caller of an Action, the same one the web page and MCP call. Processors react to events; schedules and routines run work on a timer.^[1]^[3]

AI SDK:generateText and streamText with tools, multi-step with stopWhen, and the ToolLoopAgent class. Durable workflows come from a separate package, the Workflow DevKit.^[6]^[7]^[8]

Memory

Copilotz:The history plugin summarizes long chats; the memory plugin remembers across them, by meaning with an embeddings model, or by words without one.^[3]^[9]

AI SDK:Not kept by the SDK: you store messages yourself, as its persistence guide shows. Long-term memory comes through providers such as Mem0 or Letta, or a tool of your own.^[10]^[11]

MCP (client and server)

Copilotz:Server: the mcp plugin offers the app's tools to ChatGPT, Claude and any MCP client, and draws its web Views inside them as widgets. No MCP sign-in yet: each conversation comes in as a contact. Client: not offered; the docs don't use outside MCP servers as tools.^[3]^[12]^[13]

AI SDK:Client: createMCPClient (@ai-sdk/mcp), over HTTP, SSE or stdio. Server: not in the SDK; Vercel's separate mcp-handler package hosts one.^[14]^[15]

Auth and access rules

Copilotz:Built in, under MIT. Every call names who acts, for whom and where. Each Action says who may call it (allow), asked of an agent and of whom it works for. Rules at the data filter every read and check every write, whatever the way in.^[1]^[3]

AI SDK:No built-in users, roles or data rules: those are your code. Tools can ask for approval before they run, and an optional package moves approval rules into Open Policy Agent.^[6]^[10]^[16]

Persistence

Copilotz:One store for everything: in memory, SQLite or PostgreSQL, held to one conformance suite. Moving between them changes one line.^[17]

AI SDK:None built in: you bring the database.^[10]

Observability

Copilotz:Every write and Action call is an event naming who acted, for whom and where. Metrics follow OpenTelemetry's model, exported over OTLP or as Prometheus text.^[3]^[18]

AI SDK:OpenTelemetry, through the @ai-sdk/otel package.^[19]

Testing

Copilotz:The whole app, tested offline with a scripted model and a manual clock, and no API key. copilotz/testing is one test API for Node, Deno and Bun.^[1]^[3]

AI SDK:Mock models and helpers in ai/test (MockLanguageModelV4, for one) and simulateReadableStream, to test without calling a provider.^[20]

Deployment

Copilotz:You host it: a Node web host and a store from the environment, SQLite or PostgreSQL, with a lease so a deploy hands over without two writers.^[21]

AI SDK:A library: it runs inside your app, on Vercel or any other host.^[5]

Channels

Copilotz:A web page, a terminal, a REST API and client, MCP with widgets, WhatsApp (Meta's Cloud API), voice calls (Twilio) and a chat on your site.^[3]

AI SDK:None in the SDK. Vercel's separate Chat SDK builds bots for Slack, Teams, Google Chat, Discord, Telegram and more.^[22]

License

Copilotz:MIT.^[3]

AI SDK:Apache 2.0.^[5]

Maturity

Copilotz:A release candidate: 1.0.0-rc.45, on 12 November 2026. 1.0 isn't out yet. A young project with a small community.^[23]

AI SDK:Stable: version 7 since 25 June 2026; ai 7.0.171 on 12 November 2026. About 27,900 GitHub stars and 44.8 million npm downloads a week.^[24]^[25]^[26]^[27]

Ecosystem

Copilotz:Model providers through four protocols (OpenAI, OpenAI-compatible, Anthropic and Gemini). Its own plugins for sign-in, spaces, memory, skills, routines, boards, documents and billing. Few outside integrations so far.^[3]^[28]

AI SDK:Dozens of official and community model providers, the AI Gateway as the default route, and chat hooks for React, Vue, Svelte and Angular.^[26]^[29]^[30]

import { z } from 'zod';import { generateText, tool, isStepCount } from 'ai'; const result = await generateText({ model: "anthropic/claude-sonnet-5.5", tools: { weather: tool({ description: 'Get the weather in a location', inputSchema: z.object({ location: z.string().describe('The location to get the weather for'), }), execute: async ({ location }) => ({ location, temperature: 72 + Math.floor(Math.random() * 21) - 10, }), }), }, stopWhen: isStepCount(5), prompt: 'What is the weather in San Francisco?',});

// From the README's salon.ts: one Action, three callers (excerpt).actions: { "booking.make": defineAction({ input: booking, // Asked of every caller, tools included: people, agents working for them, and contacts. allow: peopleOrContacts, run: (ctx, input) => ctx.db.collection("booking").insertOne(input), }),},views: { // What a model reads after it books, the app's own or one over MCP. "tool:book": (_, ctx) => ctx.show("bookings", {}),},resources: { // The rules: a booking lives in a space, and belongs to the chat it was made in. shared: { booking: { part: "thread" } }, tools: { book: tool({ description: "Books a slot, like 'Wed 15:00'.", input: { type: "object", properties: { slot: { type: "string" } }, required: ["slot"] }, run: ({ slot }, ctx) => ctx.call("booking.make", { slot, in: ctx.in, thread: ctx.thread }), }), }, agents: { concierge: { instructions: "Book slots.", model: "openai/gpt-5.4-mini", tools: ["book"] }, }, // An MCP server at /mcp/salon. No sign-in: each conversation is a contact, in its own chat. mcp: { salon: { in: "desk", tools: { book: {} } } },},