Build provider-agnostic AI agents and workflows in TypeScript.

Anvia is a TypeScript runtime for agents, tools, structured extraction, retrieval, pipelines, and observability inside your application code.

It gives teams more structure than raw model calls without forcing a heavyweight orchestration stack. You bring the product, data, permissions, persistence, deployment, and side effects. Anvia gives you small, explicit AI runtime contracts that fit around them.

The core design is dependency-injection oriented: your app creates provider models, typed tools, memory stores, vector indexes, observers, services, and transports, then passes the relevant objects into Anvia agents, runners, or adapters. Anvia runs the model/tool loop; your application keeps ownership of product architecture.

• Provider-neutral clients for OpenAI-compatible APIs, Anthropic, Gemini, and Mistral.
• Agent and tool APIs that keep application behavior explicit and typed.
• Embeddable runtime contracts that keep provider, memory, observability, storage, and service choices in application code.
• Structured extraction and output schemas for turning model responses into usable data.
• Pipeline primitives for composing functions, agents, extractors, batches, and parallel branches.
• Retrieval adapters for in-memory search, local embeddings, ChromaDB, Qdrant, and pgvector.
• Optional Studio, MCP, local skills, Langfuse, and OpenTelemetry integrations.

Install the core runtime and a provider adapter:

pnpm add @anvia/core @anvia/openai

Create a provider client, build an agent, and run it from your app:

import { AgentBuilder } from "@anvia/core";
import { OpenAIClient } from "@anvia/openai";

const client = new OpenAIClient({ apiKey });
const model = client.completionModel("gpt-5.5");

const supportAgent = new AgentBuilder("support", model)
  .instructions("Answer support questions clearly. Ask for missing details.")
  .build();

const response = await supportAgent
  .prompt("A customer cannot reset their password. What should I check first?")
  .send();

console.log(response.output);

Use the same runtime shape with other providers:

pnpm add @anvia/anthropic @anvia/gemini @anvia/mistral

Anvia clients take explicit constructor values and do not read environment variables on their own, so credentials stay in your existing configuration layer.

Anvia includes @anvia/studio, a local browser UI for inspecting and running agents, tools, sessions, traces, pipelines, memory, status, and knowledge during development. Add one line to serve any agent in Studio:

new Studio([agent]).start({ port: 4021 });

The cookbook is the fastest way to see Anvia in motion. It walks from a first text call through tools, structured output, providers, multimodal inputs, pipelines, retrieval, multi-agent workflows, evals, Studio, and integrations.

Run the first example from the repository root:

pnpm install
pnpm cookbook:basics:01

Run Studio locally:

pnpm cookbook:studio:01

• Docs Overview
• Install Packages
• Build your first agent
• Examples
• Packages
• Contributing

MIT