AStack

Website · Docs · Examples · Changelog · Contributing

English · 简体中文

Public beta. APIs can change between beta releases. Each release is described in the package changelogs.

AStack is a TypeScript framework for AI applications. It has one rule: everything is a component. Models, tools, agents and your own logic are components. A pipeline connects their ports into a graph, and data flows along the connections.

import { Pipeline } from '@astack-tech/core';

// Trim and Uppercase are two small components; see examples/pipeline.
const pipeline = new Pipeline();
pipeline.addComponent('trim', new Trim());
pipeline.addComponent('upper', new Uppercase());
pipeline.connect('trim.out', 'upper.in');

await pipeline.run('trim.in', '  hello  '); // 'HELLO'

The computation model

Idea	What it is	API
Component	A unit of work with named input and output ports	class X extends Component
Port	A channel. Output ports connect to input ports.	$i('in').receive(), $o('out').send()
Pipeline	A graph of components that runs data from a trigger port to an endpoint	addComponent, connect, run

A component never calls another component. It reads its input ports and writes its output ports, and the pipeline decides where data goes. Each component stays small and testable, and you can rewire a graph without editing the components in it.

Every component works in two modes: call run(input) like a function, or add it to a pipeline, where the runtime calls _transform($i, $o). Agents, multi-step workflows and routing are patterns built from these ideas, not separate subsystems.

The runtime is HLang, a flow-based programming runtime built on RxJS. Each connection is an observable stream.

How AStack compares

AStack's computation model is original. PocketFlow reached a similar idea independently: a small core in which the graph is the framework. PocketFlow is written in Python and makes a 100-line core its defining goal. AStack does not optimize for line count; it optimizes for typed, reactive data flow in TypeScript. Haystack inspired AStack's API shape. AStack does not use Haystack code.

	AStack	PocketFlow	Haystack
Language	TypeScript	Python	Python
Unit of work	Component with named ports	Node with prep, exec and post	Component with typed sockets
How data moves	Along each connection, from an output port to an input port	Through a shared store that nodes read and write	Along connections between sockets
Branching	A component sends to the output port it chooses	post returns an action; the flow follows that labeled edge	Router components
Execution	Reactive streams; concurrent runs of a route resolve in order	A flow runs nodes one after another; async and parallel variants exist	A pipeline run loop
Design goal	A typed, reactive data-flow model	A minimal core of about 100 lines	A full framework for search and RAG

The main difference from PocketFlow is where state lives. PocketFlow nodes share one store. In AStack each connection carries the data for that step, so every input and output is visible in the graph. Details: Introduction.

Install

Requires Node.js 20 or later.

npm install @astack-tech/core@beta @astack-tech/components@beta
# Model providers and tools
npm install @astack-tech/integrations@beta @astack-tech/tools@beta

An agent in a few lines

import { Agent } from '@astack-tech/components';
import { ModelProvider } from '@astack-tech/integrations';
import { createTool } from '@astack-tech/tools';

const multiply = createTool(
  'multiply',
  'Multiply two numbers.',
  args => Number(args['a']) * Number(args['b'])
);

// Any model works: see the table below.
const model = new ModelProvider.Deepseek({ apiKey: process.env['DEEPSEEK_API_KEY'] ?? '' });

const agent = new Agent({ model, tools: [multiply] });

const { message, toolCalls } = await agent.run('What is 37 times 43?');

Any model

An agent depends on one interface, chatCompletion(messages), not on a vendor. OpenAICompatible calls any API that speaks the OpenAI chat completions protocol, and Deepseek is a preset of it.

Model	Provider
OpenAI	new ModelProvider.OpenAICompatible({ apiKey, model: 'gpt-4o-mini' })
DeepSeek	new ModelProvider.Deepseek({ apiKey })
Qwen, Moonshot and other OpenAI-compatible APIs	new ModelProvider.OpenAICompatible({ apiKey, baseURL, model })
Local models through Ollama or vLLM	new ModelProvider.OpenAICompatible({ apiKey: 'ollama', baseURL: 'http://localhost:11434/v1', model: 'llama3.1' })
Any other API	An object with chatCompletion(messages): see custom-model

The examples use DeepSeek so they run as written; change the line that creates the model to use another one. Details: Model providers.

Examples

Each example is one short file that shows one idea. Read them in order; the code is also rendered on astack.tech/examples.

#	Example	Shows	API key
1	component	Define a component and run it as a function	No
2	pipeline	Connect components and infer the endpoint	No
3	routing	Branch through named output ports	No
4	text-splitter	Split documents with a built-in component	No
5	tools	Create tools and run tool calls	No
6	chat	Call a model directly and in a pipeline	Yes
7	workflow	Chain outline, draft and edit steps	Yes
8	agent	Let a model call a tool	Yes
9	custom-model	Drive an agent with your own model	No
10	streaming-agent	Stream text and tool events	Yes
11	rlm-agent	Question a codebase too large for one prompt	Yes
12	http-server	Serve a streaming agent over Server-Sent Events	Yes

pnpm install && pnpm build
pnpm --filter @astack-examples/pipeline start

Packages

Package	Contents
@astack-tech/core	Component, Pipeline, Port
@astack-tech/components	Agent, StreamingAgent, RLMAgent, TextSplitter, ToolInvoker
@astack-tech/tools	createTool, tool sets, component tools, MCP tools
@astack-tech/integrations	Model providers: OpenAICompatible for any OpenAI-compatible API, and the Deepseek preset

Each package README lists its full API.

Documentation

Page	Topic
Introduction	The computation model and the comparison above
Getting started	Install, first component, first pipeline
Components, Pipelines, Routing	The core model
Tools, Model providers	Building blocks for agents
Agents, Streaming, RLM	Agent patterns
API reference	Every export of every package, generated from the source and its TSDoc comments

Development

pnpm install
pnpm verify   # format, lint, types, tests, package checks, build

The rules for contributors and AI coding agents are in AGENTS.md: code is the single source of truth, every change starts with a failing test, and every gate runs at its strictest setting with zero warnings. Open pull requests against dev. See CONTRIBUTING.md and SECURITY.md.

License

MIT. Thanks to the HLang project for the runtime, to Haystack for the API inspiration, and to everyone who has contributed.