Schift is the TypeScript AI agent framework with built-in RAG. Scaffold, build, and deploy AI agents in minutes with templates, tool calling, and managed cloud.

How does Schift work?

Run npx create-schift to scaffold an agent project from a template. Upload documents for RAG, register tools, and deploy. The framework handles the ReAct loop, tool dispatch, and LLM routing.

How much does Schift cost?

The framework is open source and free. Schift Cloud pricing: ingestion $0.009/page, search $0.005/request, storage $0.03/GB/month, embedding $0.015/1M tokens.

Is there a free tier?

Yes. Sign up and start with free credits. No credit card required.

Ship
intelligence.

Docs in. Agent deployed.

Start with a template View docs

import { Agent, RAG } from '@schift-io/sdk'

const rag = new RAG({ bucket: 'support-docs', topK: 5 })

export const agent = new Agent({
  name: 'support-bot',
  instructions: 'Answer questions from the knowledge base.',
  rag,
  tools: [searchDocs, collectLead],
  model: 'gpt-4.1-nano',
})

const result = await agent.run('How do I reset my password?')
console.log(result.output)

import type { AgentTool } from '@schift-io/sdk'

export const searchDocs: AgentTool = {
  name: 'search_docs',
  description: 'Search the knowledge base',
  parameters: {
    type: 'object',
    properties: {
      query: { type: 'string' }
    },
  },
  handler: async ({ query }) => {
    return rag.search(query)
  },
}

$ npx create-schift@latest

  Welcome to Schift - The AI Agent Framework

? Project name: support-bot
? Template: cs-chatbot
? API key: sch_****

  Done! Next steps:
    cd support-bot
    npm run dev

TypeScript | RAG | Tools | Any LLM | Open Source

97%+

retrieval accuracy across BEIR benchmarks

<300us

p50 search latency at 1M vectors

5min

to first agent deploy

templates ready to use

Available today

Framework npx create-schift@latest TypeScript SDK npm install @schift-io/sdk Python SDK pip install schift MCP npx schift-mcp REST API api.schift.io

Primitives

Everything an agent needs. Nothing it doesn't.

Five primitives. Zero boilerplate. Your agent ships with RAG, tools, and memory out of the box.

TypeScript

Agent Runtime

ReAct loop with automatic tool dispatch. Define instructions, register tools, call agent.run(). Schift handles the rest.

agent.ts

import { Agent } from '@schift-io/sdk'

const agent = new Agent({
  name: 'support-bot',
  instructions: 'Answer from the knowledge base.',
  tools: [searchDocs],
  model: 'gpt-4.1-nano',
})

const result = await agent.run('How do I reset?')

TypeScript

Built-in RAG

Upload docs, search by meaning, get answers with citations. OCR, chunking, embedding, and reranking — all managed.

rag.ts

import { RAG } from '@schift-io/sdk'

const rag = new RAG({
  bucket: 'support-docs',
  topK: 5,
})

await rag.upload(['manual.pdf', 'faq.docx'])
const results = await rag.search('password reset')

TypeScript

Tool Calling

Register any async function as a tool. Type-safe parameters, automatic validation, error recovery. Works with OpenAI and Anthropic formats.

tools.ts

import type { AgentTool } from '@schift-io/sdk'

export const collectLead: AgentTool = {
  name: 'collect_lead',
  description: 'Collect visitor contact info',
  parameters: {
    type: 'object',
    properties: { email: { type: 'string' } },
  },
  handler: async ({ email }) => saveLead(email),
}

TypeScript

Any LLM, One Endpoint

GPT, Claude, Gemini — switch models without changing code. OpenAI-compatible routing with automatic fallback.

model.ts

const agent = new Agent({
  model: 'gpt-4.1-nano',
  // swap to 'claude-sonnet-4-6'
  // or 'gemini-2.5-flash'
  // — zero code change needed
})

// Automatic fallback if primary model fails
const result = await agent.run('Summarize this contract')

Templates

Start from a template. Ship faster.

Production-ready agent templates. Pick one, customize, deploy.

cs-chatbot

Customer Support

RAG-powered support agent with lead collection.

FAQ answering
Source citations
Lead capture

legal-qa

Legal Q&A

Contract analysis and legal document search.

Clause extraction
Case law search
Korean law support

research-qa

Research Assistant

Paper Q&A with cross-reference citations.

PDF ingestion
Cross-paper synthesis
Citation tracking

How it works

Scaffold. Build. Deploy.

Scaffold your agent

npx create-schift — pick a template, enter your API key. You get a working agent project in seconds.

$ npx create-schift@latest

Upload your data

Push documents via SDK or dashboard. Schift handles OCR, chunking, and embedding. Your agent can search them instantly.

await rag.upload(['manual.pdf', 'faq.docx'])

Deploy to production

npm run dev for local testing. schift deploy when ready. Your agent is live with built-in monitoring and logging.

$ schift deploy

Two paths

Code. Or Cloud.

Write your agent for full control, or use Schift Cloud to skip the infrastructure.

Schift SDK

npx create-schift, connect your data, deploy your agent. TypeScript-first, open source.

npx create-schift

Connect your data

Deploy your agent

Get started

Schift Cloud

Managed RAG pipeline, embedding, vector search, LLM routing, observability. Zero infrastructure.

Upload docs

Connect agent

Scale automatically

Agents for every use case.

Support "How do I reset my password?" → instant answer from your docs + source citation

Legal "What are the termination clauses?" → precise clause extraction from contracts

Research "Summarize findings on treatment X" → cross-paper synthesis with citations

Internal "What's our PTO policy?" → answer from company handbook in seconds

Live Demo

Try it yourself

Ask our AI agent anything about Schift. No signup required.

AI Schift Agent

Hi! Ask me anything about Schift — pricing, features, or how to get started.

Try asking

Performance

2.3x faster than FAISS.
3.3x better tail latency.

Same HNSW algorithm, same machine, same vectors. Schift SQ8 delivers sub-300us p50 with 3.3x tighter p99 than FAISS — at 4x less memory, with persistence, knowledge graphs, and compaction included.

1M vectors dim=1024 top-10 single-thread HNSW M=32

Schift Engine

SQ8 4x compression

p50 277us

p99 502us

QPS 3,400

FAISS HNSW

In-Memory Only

p50 621us

p99 1,653us

QPS 1,503

3.3x better tail latency, 2.3x higher throughput

Schift SQ8 p99 is 502us vs FAISS 1,653us at 1M vectors, using 4x less memory. Unlike pure in-memory libraries, Schift includes disk persistence, WAL crash recovery, knowledge graph search, and live compaction.

74K

vectors/sec projection

memory savings (SQ8)

re-embedding needed

Apple Silicon, single-thread search, same machine. HNSW M=32, efConstruction=200, efSearch=50. Full benchmark source available on GitHub.

Pricing

Simple. Transparent.

Subscription-first plans with included quota, credits for overage, and dedicated instances for larger teams.

Metric

Price

vs Competition

Ingest OCR + chunking + embedding in one price

$0.009 / page

Upstage OCR alone costs more

Search semantic search across all uploaded documents

$0.005 / request

—

Storage auto-compressed vectors and files

$0.03 / GB·month

—

Embedding direct API, 85% less than Upstage

$0.015 / 1M tokens

—

LLM OpenAI-compatible routing, one endpoint any model

cost + 5%

—

Start free See full pricing

Deploy your first agent
in 5 minutes.

Free tier included. No credit card required.

Works without a website Free 10K executions/month Deploy in 5 minutes

Get started Read the docs

Need help getting started?

Ship intelligence.

Everything an agent needs. Nothing it doesn't.

Agent Runtime

Built-in RAG

Tool Calling

Any LLM, One Endpoint

Start from a template. Ship faster.

Customer Support

Legal Q&A

Research Assistant

Scaffold. Build. Deploy.

Code. Or Cloud.

Schift SDK

Schift Cloud

Agents for every use case.

Try it yourself

2.3x faster than FAISS.3.3x better tail latency.

Simple. Transparent.

Deploy your first agent in 5 minutes.

Ship
intelligence.

2.3x faster than FAISS.
3.3x better tail latency.

Deploy your first agent
in 5 minutes.