TypeScript agent framework for building production-grade, type-safe AI applications — the Pydantic AI way, using Vercel AI SDK.

📖 vibes-sdk.a7ul.com — full documentation, guides, and API reference

Quick Start · Installation · Why Vibes · Documentation · Examples

---

Architecture

flowchart TD
    A[Your Code] --> B["@vibesjs/sdk: agent loop · tools · DI · evals · streaming · graph"]
    B --> C["Vercel AI SDK: models · streaming · structured output · providers"]
    C --> D[Anthropic]
    C --> E[OpenAI]
    C --> F[Google]
    C --> G["Groq · Mistral · Ollama · 50+ more"]

Loading

Quick Start

import { Agent } from "@vibesjs/sdk";
import { anthropic } from "@ai-sdk/anthropic";

const agent = new Agent({
  model: anthropic("claude-haiku-4-5-20251001"),
  systemPrompt: "You are a helpful assistant.",
});

const result = await agent.run("What is the capital of France?");
console.log(result.output); // "Paris"

Installation

Deno:

deno add jsr:@vibesjs/sdk

Node.js:

npx jsr add @vibesjs/sdk
npm install ai zod

Then install a provider:

# Anthropic
npm install @ai-sdk/anthropic   # ANTHROPIC_API_KEY

# OpenAI
npm install @ai-sdk/openai      # OPENAI_API_KEY

# Google
npm install @ai-sdk/google      # GOOGLE_GENERATIVE_AI_API_KEY

Vibes supports 50+ providers via the Vercel AI SDK. Switching providers is a one-line change.

AI coding assistant support

Claude Code (agent skill) — install the @vibesjs/sdk skill so your coding assistant writes idiomatic Vibes code without looking up docs:

mkdir -p .claude/agents && curl -fsSL https://raw.githubusercontent.com/a7ul/vibes/main/packages/sdk/skills/vibes-sdk.md -o .claude/agents/vibes-sdk.md

MCP server — add the Vibes docs MCP server to any MCP-compatible client (Cursor, Windsurf, Claude Desktop, etc.) for in-editor documentation access:

{
  "mcpServers": {
    "vibes-sdk": {
      "url": "https://vibes-sdk.a7ul.com/mcp"
    }
  }
}

Why Vibes?

Most AI frameworks try to hide the model behind magic. Vibes does the opposite — it gives you a thin, typed layer that stays out of the way.

The core idea is borrowed from Pydantic AI: agents are just functions. They take input, call tools, and return typed output. There's no hidden state, no opaque orchestration engine, no DSL to learn. If you know TypeScript and async/await, you already know how to use Vibes.

Feature	What it means
Type-safe tools + Dependency injection	Every tool parameter is validated at runtime with Zod. Carry databases, HTTP clients, and config via RunContext through the entire call chain. No any types, no global state.
Automatic retries + Cost control	Retries on validation failure and enforces token budgets and request limits to keep costs in check.
Structured output + Streaming	Define a Zod schema, get back a typed object — or stream typed partial objects to the client as they arrive.
Testing + Evals	Unit-test every agent in CI with TestModel and setAllowModelRequests(false) — no real API calls. Then go further with typed eval datasets, built-in and LLM-as-judge evaluators, and experiment runners with configurable concurrency. Evals are code — they live in your repo, run in CI, and catch regressions before they reach users.
Model-agnostic	Switch between Anthropic, OpenAI, Google, Groq, Mistral, Ollama, and 50+ providers by changing one line.
OpenTelemetry observability	Every run emits OTel spans, events, and token usage metrics. Works with Jaeger, Honeycomb, Datadog, and any OTel-compatible backend.
Durable agents + MCP, AG-UI, A2A	Run long-lived agents that survive crashes and restarts with Temporal. Connect to MCP servers and build AG-UI and A2A agents out of the box.

Examples

Type-safe tools + structured output

import { Agent, tool } from "@vibesjs/sdk";
import { anthropic } from "@ai-sdk/anthropic";
import { z } from "zod";

// Dependencies
type Deps = {
  db: { getUser: (id: string) => Promise<{ name: string; plan: string }> };
};

// Tools
const getUserInfo = tool({
  name: "get_user_info",
  description: "Fetch user details from the database",
  parameters: z.object({ userId: z.string() }),
  execute: async (ctx, { userId }) => ctx.deps.db.getUser(userId),
});

// Structured output schema
const SupportResponse = z.object({
  greeting: z.string(),
  recommendation: z.string(),
  escalate: z.boolean().describe("Whether to escalate to a human agent"),
});

// Agent
const supportAgent = new Agent<Deps>({
  model: anthropic("claude-haiku-4-5-20251001"),
  systemPrompt: "You are a customer support agent. Be concise and helpful.",
  tools: [getUserInfo],
  outputSchema: SupportResponse,
});

// Run with injected deps
const result = await supportAgent.run("Help user-42 with their billing question", {
  deps: {
    db: { getUser: async (id) => ({ name: "Ada", plan: "pro" }) },
  },
});

console.log(result.output.greeting); // "Hi Ada!"
console.log(result.output.escalate); // false

Testing (no API calls)

import { Agent, TestModel, setAllowModelRequests } from "@vibesjs/sdk";

setAllowModelRequests(false); // block any real LLM calls in CI

const agent = new Agent({
  model: new TestModel({ output: "Paris" }),
  systemPrompt: "You are a helpful assistant.",
});

const result = await agent.run("What is the capital of France?");
assert(result.output === "Paris");

Streaming

const stream = agent.stream("Explain async/await in TypeScript");

for await (const chunk of stream.textStream) {
  process.stdout.write(chunk);
}

Feature parity

Vibes tracks feature-for-feature parity with Pydantic AI. See the full feature parity breakdown for per-feature status and notes.

Area	Status
Agent API	✅
Tools	✅
Toolsets	✅
Deferred tools (human-in-the-loop)	✅
Output & Structured Results	✅
Message history	✅
Dependencies	✅
Usage & Limits	✅
Errors	✅
MCP (Model Context Protocol)	✅
Testing	✅
Multi-Agent & Graph	✅
Observability	✅
Evaluation framework	✅
Durable execution	✅
AG-UI Protocol	✅
Multi-Modal Support	✅

Pydantic AI comparison

Vibes is, in a real sense, Pydantic AI for TypeScript. Most concepts transfer directly.

Concept	Pydantic AI	@vibesjs/sdk
Core class	Agent	Agent
Run	agent.run_sync()	agent.run()
Streaming	agent.run_stream()	agent.stream()
Tools	@agent.tool decorator	tool() factory
Dependencies	deps: TDeps	deps: TDeps
Typed output	result_type: MyModel	outputSchema: z.object(...)
Type validation	Pydantic	Zod
Result validators	@agent.result_validator	resultValidators: [...]
Testing	TestModel / FunctionModel	TestModel / FunctionModel
Language	Python	TypeScript
Model layer	Pydantic AI providers	Vercel AI SDK

Packages

Package	Description
@vibesjs/sdk	The core agent framework

Maintained by AI agents

Vibes was created and is maintained by AI agents under the supervision of Atul (@a7ul). Every commit is reviewed by a human; every line was written by an agent.

When Pydantic AI ships a new release, a GitHub Actions workflow automatically detects it, opens an issue with a full porting checklist, and assigns the GitHub Copilot coding agent to implement it. The resulting PR is reviewed and merged by a human.

Contributing

Contributions are welcome. Please open an issue before submitting a large PR so we can discuss the approach.

flowchart LR
    A[Open Issue] --> B[Discuss Approach]
    B --> C[Fork & Branch]
    C --> D[Write Tests First]
    D --> E[Implement]
    E --> F[Open PR]
    F --> G[Review & Merge]

Loading

Acknowledgements

• Pydantic AI — the design philosophy, API shape, and abstractions that Vibes is modeled after. If you like Vibes, go star Pydantic AI.
• Vercel AI SDK — the model layer powering all providers, streaming, and structured output.
• Zod — runtime schema validation used throughout for tools, output, and dependencies.

License

MIT — see LICENSE