NeuroLink
A unified TypeScript pipe for multi-provider inference, calibrated decisions, voice, RAG, memory, and MCP tools.
- Source repo
- juspay/neurolink
- Stars
- ★ 167
- Last updated
- 1d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 56/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A TypeScript product team that wants to switch among OpenAI, Anthropic, Vertex AI, Bedrock, or local runtimes without replacing its application-facing API.
- Not a fit if
- Teams requiring a ready-made graphical interface
- Workflows that require explanations for model decisions
- Teams automating irreversible decisions without review
- Website
- neurolink.ink ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v13.13.5 (2026-11-19)
- Source review
- 56/100 · Major gaps
What does this agent do, and when should you use it?
NeuroLink is an MIT-licensed TypeScript SDK and CLI that places OpenAI, Anthropic, Google, AWS, Azure, Mistral, aggregators, and local model runtimes behind one interface. Its core API distinguishes three inference modes: generate and stream produce text, while decide returns typed boolean, choice, or score judgments with probabilities or confidence values. The repository also includes MCP connectivity, tool execution, provider fallback, RAG, conversation memory, file processing, embeddings, TTS, STT, and real-time voice interfaces. Applications invoke these capabilities through the NeuroLink class and consume completed content, asynchronous streams, structured decisions, audio, or other modality outputs; a CLI supplies terminal-facing commands. It can call hosted APIs or connect to local runtimes such as Ollama, LM Studio, and llama.cpp, while adopters remain responsible for provider credentials, model hosting, and optional storage infrastructure.
An application creates a NeuroLink instance and submits text, audio, images, or files to a selected provider. generate() returns completed output, stream() exposes an asynchronous stream, and decide() or fail-open tryDecide() maps one state plus named questions into typed boolean, choice, or score answers. Its RAG path can chunk files, create embeddings, run hybrid retrieval and reranking, and use in-memory, Chroma, PgVector, or Pinecone vector stores; the memory layer supports S3, Redis, or SQLite. The MCP subsystem connects over stdio, HTTP, SSE, or WebSocket and adds multi-server management, tool routing, caching, and batching. TTS and STT are configured on generate() or stream(), while RealtimeProcessor.connect() opens OpenAI Realtime or Gemini Live sessions. For autonomous experiments, ResearchWorker can propose changes within configured repository paths, execute a command, and extract a named metric from its output.
- A TypeScript product team that wants to switch among OpenAI, Anthropic, Vertex AI, Bedrock, or local runtimes without replacing its application-facing API.
- A support platform that routes tickets with a
choice, sorts priority with ascore, and sends uncertain results to a human reviewer. - A knowledge application that needs file chunking, embedding, retrieval, and reranking before calling
generate()orstream(). - A voice product combining multiple TTS and STT vendors or building bidirectional sessions with OpenAI Realtime or Gemini Live.
- An engineering team operating MCP tool servers that wants transport, routing, caching, batching, and multi-server control in one SDK.
- An ML researcher who wants
ResearchWorkerto modify selected files, run training commands, and evaluate experiments from a parsed metric.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Environment variables
TYPESAFE_API_KEYAI_GATEWAY_API_KEYLAYA_API_KEYLAYA_BASE_URL
How do you install or deploy this agent?
The supplied material identifies the npm package as @juspay/neurolink, but it does not show a verifiable installation command or state the minimum Node.js or package-manager version. A complete copyable installation procedure therefore cannot be established from this source alone. After installation, configure credentials for the chosen backend. For example, TypeSafe Jev uses TYPESAFE_API_KEY, or it can be reached through Vercel AI Gateway with AI_GATEWAY_API_KEY; a self-hosted Laya endpoint requires both LAYA_API_KEY and LAYA_BASE_URL.
How do you use this agent?
This source-provided example creates the shared pipe, submits text, and consumes streamed content:
import { NeuroLink } from "@juspay/neurolink";
const pipe = new NeuroLink();
const result = await pipe.stream({ input: { text: "Hello" } });
for await (const chunk of result.stream) {
if ("content" in chunk) {
process.stdout.write(chunk.content);
}
}For a typed judgment, call fail-open tryDecide(). It returns null when no decision provider is configured:
import { NeuroLink } from "@juspay/neurolink";
const pipe = new NeuroLink();
const decision = await pipe.tryDecide({
state: { ticket: "Refund request, $42, first occurrence" },
questions: {
autoApprove: {
type: "boolean",
instructions: "Approve without human review.",
},
},
});
console.log(decision?.answers.autoApprove.probability);The CLI exposes a wrapper for decide, although the supplied material does not include the CLI installation or full initialization procedure:
neurolink decide [state]What are this agent's strengths and limitations?
- Treats generation, streaming, and calibrated decisions as distinct inference types, with supported
inferenceKindsdeclared per provider. - Supports numerous cloud providers and aggregators plus Ollama, LM Studio, and llama.cpp local runtimes, reducing dependence on one model vendor.
- Its MCP layer covers stdio, HTTP, SSE, and WebSocket, with tool routing, caching, batching, and multi-server management.
- RAG, memory, file processing, embeddings, voice, and multi-provider fallback are available through the same SDK surface.
- AI-driven routing, context compaction, and tool selection are designed to fail open when no decision provider is configured.
decideis labelednextin the supplied material, creating pre-release API and behavior risk.- The documented decision models are about 68% accurate and provide no rationale, so irreversible unattended decisions are a poor fit.
- Advanced deployments can require additional provider accounts, MCP servers, Redis, S3, SQLite, or vector-store infrastructure.
- Decision-provider context limits differ sharply: roughly 768 tokens for Laya's default checkpoint versus about 33,000 for TypeSafe.
- The supplied material omits a verifiable installation command, minimum runtime version, and complete first-run setup.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| NeuroLink This agent | 56 · Major gaps | Library / SDKFree | ★ 167 | 1d ago | TypeScript | OpenAI API · Claude API |
| FlowCraft | 75 · Good | CLIFree + model costs | ★ 410 | 1d ago | Go | OpenAI API · Claude API |
| AgentScope 2.0 | 61 · Some gaps | Library / SDKFree + model costs | ★ 36k | 1d ago | Python | OpenAI API · Claude API |
| Qwen Code | 73 · Some gaps | CLIFree + model costs | ★ 31k | today | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The attachment security guide describes local path, URL, redirect, network-address, and size controls, and discloses that without allowedRoots the process may attach any readable file. The CLI gates project MCP configuration behind interactive or explicit trust. Environment variables, remote services, MCP, and AutoResearch can have external effects whose impact and recovery are only partly documented. Dependency-security evidence is limited; publisher identity is unverified and earns no credit.
The source and guides describe structured-output truncation, provider retries, failover, and concrete error paths. However, the README overview and configuration/version statements across guides cannot all be checked against this excerpt, and dependency availability is not evidenced provider by provider, so points are deducted.
The README, installation guide, and multi-agent configuration cover CLI, SDK, framework integrations, providers, tools, and deployment scenarios. Attachment policy, tool selection, and capability declarations provide boundaries. Intended audiences and excluded scenarios are not consolidated, and some capabilities are asserted rather than demonstrated, so this does not earn full marks.
The materials provide navigable documentation, installation steps, examples, troubleshooting and FAQ paths, and stated limitations; the license is explicitly MIT. The README version table and security policy provide update information, but this sample gives limited detail on maintenance ownership, support channels, and update commitments, and some examples need checking against actual configuration.
The shared interface for generation, streaming, typed decisions, MCP, voice, RAG, and providers shows practical integration value, with examples that help callers get started. Cost and latency claims are project assertions without verifiable measurement detail in this sample, while the broad feature set also brings configuration and operational overhead.
The security policy and CLI source corroborate the MCP trust flow; the attachment guide gives concrete settings and behaviors, and the README links to focused documentation. Some performance, coverage, and product-scope figures remain README claims that this sample cannot independently verify. This review uses only the supplied materials and makes no execution or independent-testing claim.
- If paths can come from untrusted callers, configure attachments.allowedRoots; the default permits reading any file accessible to the process.
- The AutoResearch example edits code, commits changes, runs commands, and may revert them. Restrict mutable paths, inspect the command and metric settings, and confirm repository state and recovery procedures first.
- The sample does not include a dependency inventory, audit results, or complete implementations, so dependency risk and all security claims cannot be verified from it.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
[13.13.5](https://github.com/juspay/neurolink/compare/v13.13.4...v13.13.5) (2026-11-19) Bug Fixes * (hitl): add opt-in confirmDirectToolCalls for direct executeTool built-in calls ([5080ef7](https://github.com/juspay/neurolink/commit/5080ef74f67994c85bb774599aed314220880bf6))View on GitHub ↗ -
Release notes excerpt
[13.13.4](https://github.com/juspay/neurolink/compare/v13.13.3...v13.13.4) (2026-11-19) Bug Fixes * (tool-routing): enforce toolRouting.timeoutMs on the decide call ([832867b](https://github.com/juspay/neurolink/commit/832867ba92f11aee1bed721d6d8f0a366c20ded1))View on GitHub ↗ -
Release notes excerpt
[13.13.3](https://github.com/juspay/neurolink/compare/v13.13.2...v13.13.3) (2026-11-19) Bug Fixes * (telemetry): redact decide error bodies before they reach spans ([ad71204](https://github.com/juspay/neurolink/commit/ad71204072a15eb6b2304836443477785b14fecb))View on GitHub ↗
FAQ
Does NeuroLink require a paid cloud model?
What happens when a decision or model provider is unavailable?
tryDecide() returns null when no decision provider is configured. Internal routing, compaction, and tool-selection optimizations are described as fail-open, while providerFallback and modelChain can retry substitutes after ModelAccessDeniedError.Can decide act as an unattended final approver?
Must RAG data be stored on one vendor's platform?
Can credentials and models vary per request?
Is NeuroLink open source?
Is NeuroLink safe to use?
Is NeuroLink still maintained?
Who makes NeuroLink?
What is NeuroLink's official website?
Thanks — we use this to improve the page.