createNeuron()

Creates a Neuron instance — the main interface for in-browser LLM chat.

import { createNeuron } from '@agent-layer-zero/dendrite'

const neuron = createNeuron(config)

Model loading starts immediately on creation. Use onProgress to track download progress.

Config Options

Required

Option	Type	Description
modelId	string	WebLLM model ID (e.g. 'gemma-2-2b-it-q4f16_1-MLC')

Performance (optional)

Option	Type	Description
worker	Worker | () => Worker	Optional Web Worker for off-thread performance. See Worker Setup. If omitted, runs on main thread.

Personality (use one)

Option	Type	Description
systemPrompt	string	Simple system prompt string
personalityDocs	PersonalityDoc[]	Typed personality documents

Generation

Option	Type	Default	Description
temperature	number	0.2	Randomness (0 = deterministic, 2 = creative)
maxTokens	number	2048	Max tokens to generate per response
frequencyPenalty	number	0.5	Penalty for repeated tokens (reduces loops)
maxHistoryTurns	number	10	How many conversation turns to include

Caching

Option	Type	Default	Description
useIndexedDBCache	boolean	true	Cache model weights in IndexedDB

Inference Mode

By default Dendrite auto-selects the best engine for the device (WebGPU → WASM → API). Override this with inference or engine.

Option	Type	Default	Description
inference	'auto' | 'fastest' | 'fallback' | 'api'	'auto'	Selection strategy. 'auto' tries WebGPU then WASM then API. 'fastest' requires WebGPU, throws otherwise. 'fallback' forces Transformers.js (widest compatibility). 'api' requires apiUrl.
engine	'webllm' | 'transformers' | 'api'	—	Force a specific engine, overriding inference.

Callbacks

Option	Type	Description
onProgress	(percent: number, text: string) => void	Model download progress (0-100)
onError	(error: Error) => void	Any error (load, generation, etc.)
onLoadingChange	(loading: boolean) => void	Loading state transitions
onGeneratingChange	(generating: boolean) => void	Generating state transitions
onInferenceSelected	(tier: { engine, mode, reason }) => void	Fires when the engine tier is chosen. engine is 'webllm' | 'transformers' | 'api', mode is 'webgpu' | 'wasm' | 'cloud'.

API Connection (optional)

Option	Type	Description
apiUrl	string	AgentLayerZero API URL
username	string	Profile username
instanceSlug	string	Instance slug
apiToken	string	Auth token for private instances

Returns

A Neuron instance.

Example

const neuron = createNeuron({
  modelId: 'gemma-2-2b-it-q4f16_1-MLC',
  personalityDocs: [
    { type: 'zero-shot', content: 'You are a pirate.' },
    { type: 'knowledge', content: 'You sailed the seven seas.' },
  ],
  temperature: 0.7,
  onProgress: (pct, text) => updateProgressBar(pct),
  onLoadingChange: (loading) => toggleSpinner(loading),
  onError: (err) => showError(err.message),
})