Nestor - An API Hub for AI Agents

Nestor is an API hub designed to empower AI agents with the ability to discover, aggregate, and utilize various API services available on the local network. By providing a unified interface for AI models to access a wide range of services, Nestor enables the creation of ubiquitous AI agents that can spontaneously interact with their environment and automate tasks.

Key Features

• Zeroconf Service Discovery: Nestor automatically detects and aggregates Nestor-compatible API services on the local network, making them accessible to connected AI agents. No configuration needed!
• Execution Orchestration: Nestor acts as a single point of contact for these API services, orchestrating function calls requested by the LLM model

High-Level Architecture

A typical Nestor architecture comprises of:

• A Hub that acts as the collector and orchestrator of all Nestor services
• A number of Nestor services. They can be embedded in existing software or be built as independent software who act as a proxy to other API services
• A number of Nestor clients: typically LLM-Based AI Agents that are going to expose to the LLM model all the API visible from the Hub

Getting Started

The simplest way to get started is to see Nestor in action using the example services and clients available in nestor-examples. You can watch a demo video or another one of Nestor.

You will need four different terminal windows:

1. In nestor, go to the hub directory then run

npm install
npm start

2. In nestor-examples, go to the service1 directory then run

npm install
npm start

You should see the Hub acknowledging the discovery of 2 new endpoints

3. In nestor-examples, go to the service2 directory then run

npm install
npm start

You should see the Hub acknowledging the discovery of 3 new endpoints

4. In nestor-examples, go to the client directory then run

npm install
OPENAI_API_KEY=xxxx npm start

Replace xxx with your OpenAI API Key.

The client is going to access the hub and expose the 5 enpoints to the LLM. You should see each service being called and the AI agent in action to create a playlist of the songs played at The National concert (this is of course a fake setlist). A more detailed description of what is going on is avaiable in the nestor-examples README.md.

Running the hub

Simply clone the repo, go to the hub folder and start it.

A pm2 ecosystem file is provided if you want to run it that way. Just use npm run pm2.

The hub automatically broadcasts its presence on the network to be discoverable by services and clients.

Service implementation

Implementing a service is pretty straightforward:

• Add nestor-service as a new dependency
• Import/require NestorService in your application
• Instantiate a new NestorService instance by prodiving:
  • A unique name for your service
  • The port number your service runs on
  • The URL path that lists all the API of your service
• Start the service: it should auto-discover the hub and register to it

The URL path that lists all API should return a list of API endpoints in the format { endpoints: [...] }, each endpoint being described by:

Attribute	Type	Description
name	string	Name of your API. This must be unique in your service.
description	string	The description of your API as it will be provided to the LLM. Remember that the LLM decision to call or not your API will be based on that description!
url	string	The full URL of your API including host and port. See the examples how to automatically build this from Express Request object
method	string	The HTTP method to use. Defaults to GET.
parameters	Array	List of parameters required to call your API. Check table below for the parameters attributes

Each parameter is described by:

Attribute	Type	Description
name	string	Name of the parameter
type	string	Types supported are defined by the LLM engine
description	string	The description of the parameter as it will be provided to the LLM. Remember that the LLM decision how to use the parameter will be based on that description!
required	boolean	If this paramter is mandatory or not

In specific environments, you may want to not use auto-discovery and manual register to a hub knowing its hostname and port. In that case, you can pass { autostart: false } when building the service and then use the register and unregister methods of NestorService.

Services can of course be implemented in any language. You can find an example of python service in nest-examples/service3.

Client implementation

Implementing a client is even more straightforward if you already have a LLM-based agent. If not the implementation available in nestor-examples is pretty generic (except for the hard-coded prompt) and could be reused almost as-is:

• Add nestor-client as a new dependency
• Import/require NestorClient in your application
• Instantiate a new NestorClient
• Provide the result of NestorClient.list() to the LLM call
• Call NestorClient.call() when the LLM invokes a tool

In specific environments, you may want to not use auto-discovery and manual connect to a hub knowing its hostname and port. In that case, you can pass { autostart: false } when building the client and then use the connect and disconnect methods of NestorClient.

Check the client implementation in nestor-examples.

Suggestions

Nestor is in its early stage and all suggestions are welcome!

Applications using Nestor

Clients

• Witsy: desktop AI assistant

Services

• TIDAL Streamer: Wrapper for TIDAL (music service) APIs

TODO

• Tests!
• Error handling (everywhere)
• Support more HTTP verbs for API calls
• Publish @nestor/* nestor-* packages