# NeuroLink SDK — Complete Reference

> NeuroLink is an open-source TypeScript AI SDK providing unified access to 13+ AI providers through a single consistent API.

Generated: 2026-02-24
Summary version: https://neurolink.ink/llms.txt
Documentation: https://docs.neurolink.ink

---

## Installation

npm install @juspay/neurolink
# or
pnpm add @juspay/neurolink
# or use CLI without installing
npx @juspay/neurolink generate "Hello"

## Quick Start (SDK)

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  prompt: "Explain quantum computing",
  provider: "openai",
  model: "gpt-4o",
});

console.log(result.text);

## Quick Start (CLI)

# Set provider API key
export OPENAI_API_KEY="sk-..."

# Generate text
neurolink generate "Explain quantum computing"

# Stream text
neurolink stream "Write a poem about TypeScript"

# Interactive session
neurolink loop

## Supported Providers (13+)

| Provider | Slug | Default Model | Auth |
|----------|------|---------------|------|
| OpenAI | openai | gpt-4o | OPENAI_API_KEY |
| Anthropic | anthropic | claude-sonnet-4-20250514 | ANTHROPIC_API_KEY |
| Google AI Studio | google-ai | gemini-2.5-flash | GOOGLE_AI_API_KEY |
| Google Vertex AI | vertex | gemini-2.5-flash | GOOGLE_CLOUD_PROJECT + ADC |
| AWS Bedrock | bedrock | anthropic.claude-3-5-sonnet | AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY |
| Azure OpenAI | azure | gpt-4o | AZURE_OPENAI_API_KEY + AZURE_OPENAI_ENDPOINT |
| Mistral | mistral | mistral-large-latest | MISTRAL_API_KEY |
| LiteLLM | litellm | gpt-4o (via proxy) | LITELLM_API_KEY |
| Ollama | ollama | llama3.1 | (local, no key needed) |
| Hugging Face | huggingface | meta-llama/Llama-3.1-8B-Instruct | HUGGING_FACE_TOKEN |
| AWS SageMaker | sagemaker | (custom endpoint) | AWS credentials |
| OpenRouter | openrouter | openai/gpt-4o | OPENROUTER_API_KEY |
| OpenAI-Compatible | openai-compatible | (configurable) | OPENAI_COMPATIBLE_API_KEY |

## Core SDK Methods

### neurolink.generate(options)
Generate a text response from any provider.

Parameters:
- prompt: string — The input prompt
- provider?: string — Provider slug (openai, anthropic, google-ai, etc.)
- model?: string — Model name
- maxTokens?: number — Max response tokens
- temperature?: number — Creativity (0-1)
- systemPrompt?: string — System instruction
- thinkingLevel?: "minimal" | "low" | "medium" | "high" — Extended thinking
- structuredOutput?: ZodSchema — Zod schema for typed JSON output
- rag?: RAGConfig — RAG pipeline configuration
- input?: { text, images, files } — Multimodal input

Returns: { text, usage, toolCalls, evaluation }

### neurolink.stream(options)
Stream a text response with real-time events.

Same parameters as generate(). Returns an async iterable of stream events.

24 event types: text-delta, tool-call, tool-result, finish, error, etc.

### neurolink.addExternalMCPServer(name, config)
Add an external MCP server.

Transports: stdio, http, sse, websocket

Example (stdio):
await neurolink.addExternalMCPServer("GitHub", {
  command: "npx",
  args: ["-y", "@modelcontextprotocol/server-github"],
  transport: "stdio",
  env: { GITHUB_TOKEN: process.env.GITHUB_TOKEN }
});

Example (HTTP):
await neurolink.addExternalMCPServer("remote-server", {
  transport: "http",
  url: "https://api.example.com/mcp",
  headers: { Authorization: "Bearer TOKEN" },
  timeout: 15000,
  retries: 5,
});

### neurolink.registerTool(tool)
Register a custom tool for AI function calling.

## RAG (Retrieval-Augmented Generation)

### Simple API
const result = await neurolink.generate({
  prompt: "What are the key points?",
  rag: {
    files: ["./docs/guide.md", "./data/faq.json"],
    strategy: "markdown",
    chunkSize: 512,
    chunkOverlap: 50,
    topK: 5,
  }
});

### Chunking Strategies (9)
character, recursive, sentence, token, markdown, html, json, latex, semantic

### Vector Stores (22)
In-memory, Pinecone, Weaviate, Qdrant, Chroma, Milvus, and 16 more

### Rerankers (5)
simple, llm, batch, cross-encoder, cohere

### Hybrid Search
BM25 lexical + vector similarity with reciprocal rank fusion or linear combination

## MCP Integration

58+ external servers supported:
GitHub, Slack, PostgreSQL, Redis, Google Drive, Google Maps, Brave Search, Puppeteer, Filesystem, SQLite, Memory, Fetch, Sequential Thinking, and 45+ more

4 transport types: stdio, HTTP/Streamable HTTP, SSE, WebSocket

MCP Enhancements (unique to NeuroLink):
- ToolRouter: Intelligent tool routing across multiple servers
- ToolCache: Response caching for repeated tool calls
- RequestBatcher: Batch multiple tool requests for efficiency

## Multi-Agent Workflows

3 topologies:
- Hub-Spoke: Coordinator delegates to specialist agents
- Mesh: Peer-to-peer agent collaboration
- Pipeline: Sequential processing chain

AgentNetwork manages lifecycle, routing, and shared context.

## Memory System (3 layers)

1. Conversation History: Full message history (Redis or in-memory)
2. Semantic Recall: Vector-based memory for relevant past context
3. Working Memory: Short-term state for current task

## Voice Processing

8 TTS providers, 6 STT providers
Speech-to-speech pipeline via Gemini Live integration

## Observability

- Langfuse: Token tracking, cost analysis, trace visualization
- OpenTelemetry: Distributed tracing (Jaeger, Zipkin compatible)
- 14 evaluation scorers
- External TracerProvider support

## Context Compaction (4-stage pipeline)

1. Tool output pruning
2. File read deduplication
3. LLM summarization (9-section structured summary)
4. Sliding window truncation

Auto-triggered at 80% context usage via BudgetChecker.

## CLI Commands

neurolink generate <prompt> — Generate text
neurolink stream <prompt> — Stream text
neurolink loop — Interactive REPL session
neurolink setup — Configure providers
neurolink status — Check provider health
neurolink mcp list — List MCP tools
neurolink config — Manage configuration

## Links

- npm: https://www.npmjs.com/package/@juspay/neurolink
- GitHub: https://github.com/juspay/neurolink
- Docs: https://docs.neurolink.ink
- Blog: https://blog.neurolink.ink
- License: MIT