Agent Smith

An api to create local first human friendly agents in the browser or Nodejs

📚 Read the documentation

• The body
  • Overview
  • Install
  • Basic agent
  • Interactions
    • Talk
    • Components
    • Confirm
• The brain
  • Overview
  • Install
  • Backends
  • Experts
  • Brain
  • Grammars
  • Browser
  • Templates
    • Basics
    • History
    • Few shots
• Jobs
  • Get started
  • Create a job
  • Config
  • State management
  • Memory
• Lm task
  • Get started
  • Specification
  • Use tasks
  • Variables
  • Templates
• Transient memory
  • Get started
  • Usage
  • Api
• Semantic memory
  • Get started
  • Initialize
  • Write operations
  • Read operations
• Terminal client
  • Install
  • Overview
  • Config
  • Tasks
  • Actions
  • Workflows
  • Commands
  • Options
  • Plugins
    • Overview
    • Vision
    • Code
      • Git
    • Web
      • Video
• Server
  • Get started
  • Configuration
  • Tasks
  • Endpoints
• Examples
  • Data viz

Check the 💻 examples

What is an agent?

An agent is an anthropomorphic representation of a bot. It can:

• Think: use language model servers to perform inference queries
• Interact: perform interactions with the user and get input and feedback
• Work: manage long running jobs with multiple tasks, use custom terminal commands
• Remember: use transient or semantic memory to store data

Packages

Version	Name	Description	Nodejs	Browser
	@agent-smith/body	The body	❌	✅
	@agent-smith/brain	The brain	✅	✅
	@agent-smith/jobs	Jobs	✅	✅
	@agent-smith/tmem	Transient memory	❌	✅
	@agent-smith/tmem-jobs	Jobs transient memory	❌	✅
	@agent-smith/smem	Semantic memory	✅	❌
	@agent-smith/tfm	Templates for models	✅	✅
	@agent-smith/lmtask	Yaml model task	✅	✅
	@agent-smith/cli	Terminal client	✅	❌

Philosophy

• Composable: the packages have limited responsibilities and can work together
• Declarative: focus on the business logic by expressing features simply
• Explicit: keep it simple and under user control: no hidden magic

FAQ

• What local or remote inference servers can I use?

Actually it works with Llama.cpp, Koboldcpp and Ollama.

It also works in the browser using gpu only inference and small models

• Can I use this with OpenAI or other big apis?

Sorry no: this library favours local first or private remote inference servers

Example

Terminal client

Simple inference query (using the inference plugin):

lm q list the planets of the solar system

Query a thinking model, use qwq (from the models plugin)::

lm think "solve this math problem: ..." m=qwq

Compare images (using the vision plugin):

lm vision img1.jpg img2.jpg "Compare the images"

Generate a commit message in a git repository (using the git plugin):

lm commit

Terminal client plugins

Version	Name	Description	Doc
	@agent-smith/feat-models	Models	doc
	@agent-smith/feat-inference	Inference	doc
	@agent-smith/feat-vision	Vision	doc
	@agent-smith/feat-git	Git	doc

Nodejs example

const backend = useLmBackend({
    name: "koboldcpp",
    localLm: "koboldcpp",
    onToken: (t) => process.stdout.write(t),
});

const ex = useLmExpert({
    name: "koboldcpp",
    backend: backend,
    template: templateName,
    model: { name: modelName, ctx: 2048 },
});
const brain = useAgentBrain([expert]);

console.log("Auto discovering brain backend ...");
await brain.init();
brain.ex.checkStatus();
if (brain.ex.state.get().status != "ready") {
        throw new Error("The expert's backend is not ready")
    }
// run an inference query
const _prompt = "list the planets of the solar sytem";
await brain.think(_prompt, { 
   temperature: 0.2, 
   min_p: 0.05 
});

Server api example

To execute a task using the server http api:

import { useServer } from "@agent-smith/apicli";

const api = useServer({
    apiKey: "server_api_key",
    onToken: (t) => {
        // handle the streamed tokens here
        process.stdout.write(t)
    }
});
await api.executeTask(
    "translate", 
    "Which is the largest planet of the solar system?", 
    { lang: "german" }
)

Libraries

The cli is powered by:

• Nanostores for the state management and reactive variables
• Locallm for the inference api servers management
• Modprompt for the prompt templates management

The server is powered by:

• Echo