Skip to content

NeuralCleave v2.1.0

Choose a tag to compare

View all tags
@TheAmitChandra TheAmitChandra released this 22 Jul 19:11
app-v2.1.0
2e6aabc

Full OpenClaw parity achieved. NeuralCleave now leads in 13 capability categories.

Install / Upgrade

pip install --upgrade neuralcleave

One-liner (no Python knowledge required):

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

# Windows (PowerShell)
iwr -useb https://neuralcleave.com/install.ps1 | iex

What's New in v2.1.0

📡 Channels — 32 total (up from 12)

20 new channel adapters added, covering all major platforms:

Adapter Protocol
Line HMAC-SHA256 webhook + push-message API
Feishu / Lark v1 + v2 event schema; tenant access token cache
Zalo HMAC-SHA256 webhook; OAuth2 refresh token with auto-renewal
WeChat Work SHA1 URL challenge; XML inbound; access token cache
QQ Bot Ed25519 webhook; op=13 challenge; AT/C2C/GROUP/DIRECT events
Tlon / Urbit Eyre HTTP API; SSE channel; legacy + modern letter formats
iMessage BlueBubbles REST polling; direct/SMS/group targets
Google Chat JWT service account OAuth2; space + thread targets
Messenger Meta Graph API v19.0; HMAC-SHA256 webhook
Rocket.Chat DDP WebSocket; REST v1 login; SHA-256 password digest
Bluesky AT Protocol XRPC polling; JWT auto-refresh; threaded replies
Viber REST + HMAC-SHA256 webhook; 7 inbound event types
XMPP slixmpp asyncio; MUC rooms (XEP-0045); SCRAM-SHA-1
Nostr NIP-04 encrypted DMs; pure-Python secp256k1 + BIP-340 Schnorr
Synology Chat SYNO.Chat.External API; outgoing webhook
Twitch IRC-over-WebSocket; IRCv3 tags; multi-channel join; auto-reconnect
Twilio Voice Multi-turn speech via TwiML; HMAC-SHA1 verification

🤖 LLM Providers — 13 total, 19 model aliases

8 new providers added, covering all major Chinese models OpenClaw supports plus more:

  • Mistral AI — mistral-large for complex reasoning
  • xAI Grok — grok-3 for complex reasoning
  • Cohere v2 — command-r-plus for summarization
  • Moonshot AI / Kimi — moonshot-v1-8k for general routing
  • Zhipu AI GLM — glm-4-flash for intent extraction + cheap inference
  • Alibaba Qwen / DashScope — qwen-max for code generation + review
  • Baidu ERNIE — ERNIE-Speed-128K for cheap inference
  • ByteDance Doubao — doubao-lite-4k for cheap inference

All providers use ENV:VAR_NAME secret resolution — no secrets in config files.

🧠 Multi-Agent Orchestration

Named sub-agents with isolated memory namespaces, task-type routing, keyword routing, channel-pattern (glob) routing, priority + round-robin tie-breaking, and per-node max_concurrent limits.

  • 7 REST endpoints under /api/v1/orchestrator/
  • neuralcleave orchestrate list/add/remove/route/status CLI
  • MemoryNamespaceStore — LRU key-value store per node with configurable eviction

🛒 Hub Marketplace

Install community plugins safely with dual-pass security scanning before anything touches your system.

  • PackageScanner — AST walk (13 blocked imports) + regex (14 dangerous patterns)
  • SHA-256 checksum verification on every download
  • neuralcleave hub list/search/install/remove/info/enable/disable/scan/status

✏️ Self-Modifying Skills

The AI can write, validate, and hot-load new Python tools in the middle of a conversation — no gateway restart required.

  • SkillWriter — AST validate + blocked-import check → persist to ~/.neuralcleave/skills/
  • WriteSkillTool / ListSkillsTool / DeleteSkillTool — LLM-callable in conversation
  • neuralcleave skills write/list/show/delete/validate

🎨 Visual Canvas (A2UI)

A live rendering surface the AI can write to during any conversation.

  • 7 block types: text, markdown, image, table, code, chart, html
  • 200-block ring buffer with real-time WebSocket broadcast
  • 9 LLM-callable actions via CanvasTool
  • Live page at /canvas; bar/line/pie charts via Canvas API
  • neuralcleave canvas open/status/clear/render

📱 Progressive Web App

Install NeuralCleave on iOS or Android from any browser — no app store required.

  • W3C Web App Manifest; Service Worker with offline support + Web Push
  • Full WebSocket chat UI with message streaming
  • Works on any device that can reach your gateway

🔒 Sandbox Execution

Three sandbox backends for tool execution: local subprocess, Docker (network-isolated, memory/CPU limited), and SSH remote.

🔑 Security

  • Optional X-API-Key gateway middleware — enforces auth on all /api/* routes when configured
  • /health and WebSocket endpoints always exempt (monitoring-friendly)

🔄 Skill Hot-Reload

Reload any plugin without restarting the gateway.

  • reload_plugin(name) / reload_all() on PluginRegistry
  • POST /api/v1/plugins/{name}/reload
  • neuralcleave plugins reload [name]

🖥️ Desktop App

  • Tauri 2.x sidecar pipeline: system tray, Ctrl+Shift+Space global hotkey, single-instance guard, close-to-tray
  • AutostartManager — Windows registry, macOS launchd, Linux systemd — neuralcleave autostart enable/disable/status
  • neuralcleave cloud generate/check/status — auto-generates Dockerfile, docker-compose, Railway, Render, Fly.io manifests

🧪 Tests

5,064 tests passing (was ~2,500 at v2.0.0). New suites cover gateway auth, orchestrator routing, all new channel adapters, Hub installer, SkillWriter, canvas, PWA, and sandbox.

Upgrading from v2.0.0

No breaking changes. New channels are opt-in via config.toml. New LLM providers require their respective API keys. All existing configs continue to work as-is.

Links

Assets 2
Loading