v2.1.5 — 5,966 tests passing

NeuralCleave

The intelligence layer for every conversation.

Your personal AI assistant that brings every AI model, conversation, memory, and workflow together in one private, extensible intelligence layer.

Get Started Explore Docs

🧠

3-Tier Memory

Redis, Qdrant & SQLite — your AI actually remembers you.

🔀

Smart Routing

10 task types auto-routed to the optimal model.

📡

32 Channels

Telegram, WhatsApp, Discord, Slack and 29 more.

🤖

13 Providers

Anthropic, OpenAI, DeepSeek, Ollama and more.

🎙

Voice Ready

Whisper STT, 3-tier TTS, wake word on all platforms.

🔌

Plugin SDK

Typed ABCs, entry-points, hot-reload, Hub marketplace.

32Channel Adapters

13LLM Providers

5,076Tests Passing

41REST Endpoints

3-tierMemory System

v2.1Current Release

Why NeuralCleave

Everything you'd want. Nothing you don't.

Built to surpass every personal AI assistant on the market — feature by feature, test by test.

🧠NeuralCleave leads

3-Tier Memory

Redis (hot session) + Qdrant (vector semantic) + SQLite (long-term). Per-agent namespace isolation. Compaction + archiver included. Works fully offline.

🔀NeuralCleave leads

Task-Aware Model Routing

10 task types automatically routed to the optimal provider — complex reasoning to Claude Opus, code to DeepSeek, cheap inference to Ollama. Privacy mode forces all calls local.

✨Unique

ReflectionEngine

Every response scored across 4 dimensions — Relevance, Completeness, Accuracy, Tone — on a 0–100 scale. Low-scoring responses are automatically regenerated once.

🎙️NeuralCleave leads

Voice Pipeline

Local Whisper STT (tiny→large-v3), 3-tier TTS (ElevenLabs → Kokoro → system), OpenWakeWord always-on detection, voice cloning. Everything runs on-device.

🧩NeuralCleave leads

Plugin SDK

Typed Python ABCs with PEP 451 entry-points. Hot-reload without restart. Hub marketplace with PackageScanner safety gate — no supply-chain risk.

🤖Parity

Multi-Agent Orchestrator

Named nodes with model overrides, task/keyword/channel routing, priority + round-robin. Per-node LRU memory namespace. Full REST + CLI management.

📊NeuralCleave leads

Prometheus Observability

13 built-in metrics — counters, gauges, histograms — at /api/v1/metrics. Structured JSON logging with ContextLogger. Pre-built Grafana dashboard included.

🖼️Parity

Live Canvas

The agent can render text, markdown, code, tables, charts, and images to a live visual block stream. Real-time WebSocket broadcast to any connected viewer.

📱Parity

Desktop + PWA

Tauri v2 native app for Windows / macOS / Linux. PWA served from the gateway — installable on iOS and Android from any browser. No app store needed.

Pipeline

How NeuralCleave works

Every message you send goes through a four-stage pipeline designed to give you the best possible response, every time.

Message arrives on any channel

You send a message from Telegram, WhatsApp, Discord, Slack, voice, or any of the 32 supported adapters. The gateway picks it up instantly.

Memory enriches the context

3-tier memory (Redis hot cache → Qdrant vector search → SQLite long-term) retrieves everything relevant — past conversations, preferences, and facts you've shared.

Smart routing picks the best model

The task router classifies your request across 10 task types — coding, reasoning, summarization, image analysis, and more — then sends it to the optimal LLM from 13 providers.

ReflectionEngine scores and replies

Before sending, the 4-dimension ReflectionEngine scores the response for quality, relevance, safety, and tone. If it fails the threshold, it auto-retries with a refined prompt.

Any Channel→Gateway→3-Tier Memory→Task Router→Best LLM→ReflectionEngine→Reply

01/ 10

Scroll to explore

Product Gallery

See it in action.

Every surface is designed for the long session — focused, dark by default, and built around how you actually work.

NeuralCleave — AI Chat

Product Gallery

See it in action.

Every surface is designed for the long session — focused, dark by default.

NeuralCleave — AI Chat

AI Chat

Multi-session conversations with a collapsible sidebar, streaming responses, multi-phase thinking indicator, and a glass-morphism input bar.

Use Cases

Built for real workflows

NeuralCleave adapts to how you actually work — not the other way around.

🌅

Daily Briefing

Your morning, automated

NeuralCleave pulls your calendar, checks the weather, skims priority emails, and sends a crisp summary to your Telegram at 8 AM — every day, no setup beyond a TOML file.

Scheduler → Telegram13 LLM summarizationCalendar + email tools

💻

Dev Assistant

Code help, wherever you code

Ask questions in Slack, get PR review comments inline, run terminal commands by voice, and keep a running context of your current project — across every conversation.

Slack + voice inputOllama for local privacyGitHub plugin SDK

🔬

Research Agent

Research with persistent memory

Ask anything across sessions. NeuralCleave remembers what you've already found, cross-references Qdrant vector memory, and saves structured notes to Notion or Obsidian.

3-tier vector memoryDeepSeek reasoning modelNotion plugin

🌍

Multi-language

Chat in any language

Switch between languages mid-conversation. NeuralCleave detects locale per channel, adapts tone, and maintains separate memory namespaces for each agent node.

32 channel adaptersPer-node namespace isolationMoonshot + Qwen + ERNIE

🏠

Home Automation

Control your environment by voice

Wake word detection triggers NeuralCleave on any platform. Ask it to set timers, play music, dim lights, or brief you on your day — hands-free, locally processed.

OpenWakeWordWhisper STT + 3-tier TTSTool plugin SDK

🛡️

Privacy First

Fully local — zero cloud required

Run the entire stack on your machine with Ollama. No API keys sent to third parties, no data leaving your device. Toggle a privacy flag in TOML and you're air-gapped.

Ollama local inferenceprivacy_mode = trueSelf-hosted Redis + Qdrant

Universal Reach

32 Channel Adapters

Every adapter produces a normalised InboundMessage — one interface for every platform. Configure any channel with a single TOML block.

✈️Telegram

🎮Discord

#️⃣Slack

💬WhatsApp

📧Email

📱SMS / Twilio

🟦Microsoft Teams

🔵Google Chat

🟩Matrix

🖥️IRC

🔒Signal

🟩LINE

🦅Feishu / Lark

🍎iMessage

💾Synology Chat

⚡Nostr

📞Twilio Voice

🇻🇳Zalo OA

💼WeChat Work

🐧QQ Bot

🌐Tlon / Urbit

💙Facebook Messenger

🚀Rocket.Chat

🟣Twitch

🦋Bluesky

💜Viber

💬XMPP / Jabber

⚓Mattermost

🐘Mastodon

☁️Nextcloud Talk

🪝Generic Webhook

🔌Built-in WS / REST

~/.neuralcleave/config.toml

[channels.telegram]
bot_token = "ENV:TELEGRAM_BOT_TOKEN"

[channels.discord]
bot_token = "ENV:DISCORD_BOT_TOKEN"

[channels.slack]
app_token = "ENV:SLACK_APP_TOKEN"

Model Agnostic

13 LLM Providers. One Interface.

Task-aware routing sends each request to the optimal model. Switch providers at runtime with a single API call. All keys use ENV: resolution — no secrets in files.

AnthropicPrimary

Claude — complex reasoning, extended thinking

Google Gemini

Summarisation, cheap inference, general tasks

OpenAI

Code review, broad fallback chain

DeepSeek

Code generation & review

Mistral AI

Complex reasoning fallback

xAI Grok

Complex reasoning

Cohere

Summarisation

Moonshot / Kimi

General tasks

Zhipu GLM

Intent extraction, cheap inference

Alibaba Qwen

Code generation

Baidu ERNIE

General tasks

ByteDance Doubao

Cheap inference

OllamaPrivacy mode

Fully local / offline — any model

Task-Aware Routing

Task Type	Primary Model
complex_reasoning	Claude Opus 4.8
code_generation	DeepSeek Coder
code_review	DeepSeek Coder
summarization	Gemini 2.5 Flash
intent_extraction	Gemini 2.5 Flash
task_decomposition	Claude Sonnet
reflection	Gemini 2.5 Flash
cheap_inference	Ollama (local)
general	Gemini 2.5 Flash

Override at runtime: POST /api/v1/settings/model

How We Compare

Built to lead, not follow.

A full-depth comparison against OpenClaw — the leading open-source personal AI assistant. NeuralCleave leads or matches in every tracked dimension.

Feature

Competitors

NeuralCleave

Language

TypeScript / Node.js

Python — better AI/ML ecosystem

Memory system

Flat files / LanceDB

Redis + Qdrant + SQLite (3-tier)

Per-agent memory isolation

❌

✅ LRU namespace per node

LLM providers

4–8

13 (all major + 5 Chinese)

Task-aware routing

❌

✅ 10 task types → optimal model

Response quality scoring

❌

✅ ReflectionEngine (4D, auto-retry)

Extended thinking

❌

✅ Anthropic budget_tokens

Prometheus metrics

❌

✅ 13 built-in metrics

Local / offline mode

Limited

✅ Full Ollama + privacy toggle

Voice STT + TTS

Mobile-only

✅ Whisper + 3-tier TTS (all platforms)

Wake word detection

macOS + iOS only

✅ OpenWakeWord (all platforms)

Plugin SDK isolation

Markdown files

✅ Typed ABCs + PEP 451 entry-points

Channel count

~29

Test coverage

~200

5,966 tests

Desktop app

✅

✅ Tauri v2 (Win / macOS / Linux)

Mobile companion

✅ Native apps

✅ PWA (no app store needed)

Agent orchestration

✅ Cross-machine

✅ Multi-node + namespace isolation

Full analysis in docs/COMPETITIVE_ANALYSIS_OPENCLAW.md

Start in 30 seconds.

One command. No account required. Runs entirely on your machine.

Linux / macOS

curl -fsSL https://neuralcleave.com/install.sh | bash

Copy

Windows (PowerShell)

iwr -useb https://neuralcleave.com/install.ps1 | iex

Copy

pip

pip install neuralcleave

Copy

Docker

docker run -p 7432:7432 ghcr.io/theamitchandra/neuralcleave:latest

Copy

Desktop App

NeuralCleave v2.1.5 — Native Tauri App

PyPI

Windows.exe Windows.msi macOS.dmg Linux.AppImage All releases →

Python 3.12+ required for CLI install. Desktop app ships as a self-contained bundle — no Python needed.

Read Getting Started →View on GitHub