{"slug": "show-hn-openvole-4-5-is-out", "title": "Show HN: OpenVole 4.5 Is Out", "summary": "OpenVole 4.5, a self-hosted agent operating system for running and managing AI agents on personal hardware, has been released. The platform is model-agnostic, supports peer-to-peer networking via VoleNet, and features a microkernel architecture with swappable plugins called Paws. It allows users to create, configure, and chat with multiple agents from a single dashboard while maintaining privacy and control.", "body_md": "**The self-hosted agent OS — run a fleet of AI agents on your own hardware,\nagainst any model, from one dashboard. Peer-to-peer networked. Nothing phones home.**\n\nOpenVole is a **self-hosted server for running and managing AI agents**. One command —\n`vole serve`\n\n— gives you a dashboard where you create, configure, and chat with a whole\nfleet of agents, each isolated with its own tools, memory, and identity.\n\n**Model-agnostic**— Gemini, OpenAI, Claude, xAI, or local Ollama. Your choice, swappable per agent.** Self-hosted & private**— runs entirely on your hardware, against whatever model you point it at, with nothing phoning home.** Networked**— connect instances over** VoleNet**, a peer-to-peer agent mesh, and share tools, memory, and even a brain across machines.\n\nUnder the hood it's a **microkernel**: the core is just the agent loop and a plugin contract.\nEvery capability — reasoning, memory, tools, channels — is a swappable plugin (a **Paw**), so\nyou're never locked into someone else's worldview. A fresh install ships with zero baked-in\nopinions: you assemble exactly the agent you want.\n\n```\nnpm install -g openvole          # package is \"openvole\"; gives you the `vole` command\nmkdir my-agents && cd my-agents\nvole serve\n```\n\n`vole serve`\n\nprints a **tokenized dashboard URL** (the dashboard is gated by a session token) —\nopen it, click **New space**, and onboarding installs the essential paws (brain, session, memory,\ncompact, shell). Point a space at your model:\n\n```\nBRAIN_PROVIDER=gemini\nGEMINI_API_KEY=your-api-key\n```\n\n…then start it and chat in the **Chat** tab. Prefer one command? Use a preset:\n\n```\ncurl -fsSL https://raw.githubusercontent.com/openvole/openvole/main/presets/basic.sh | bash\n```\n\nThe npm package is\n\n(it provides the`openvole`\n\n`vole`\n\ncommand). To run without installing:`npx openvole <command>`\n\n—not`npx vole`\n\n, which is an unrelated package.\n\nFull walkthrough → [Getting Started](https://openvole.github.io/openvole/getting-started).\n\n🏠 Self-hosted & private |\nYour agents, your hardware, your data. Nothing phones home — point it at a local Ollama model and it never touches the cloud. |\n🔌 Model-agnostic |\nOne unified brain paw speaks Anthropic, OpenAI, Gemini, xAI, Ollama, and the local Claude Code CLI. Switch providers per agent with a single env var. |\n🖥️ A server, not a script |\n`vole serve` is a control plane: create, start, stop, and chat with a fleet of agents — \"spaces\" — from one browser dashboard. No babysitting processes on ports. |\n🧩 Microkernel, zero lock-in |\nA tiny, LLM-ignorant core. Every capability is a Paw you can swap, sandbox, or write yourself — and you can bring your own system prompt via `BRAIN.md` . |\n🕸️ VoleNet |\nA peer-to-peer AI agent network: remote tools become local, cheap workers share one brain, memory syncs across the mesh — signed with hybrid post-quantum signatures (Ed25519 + ML-DSA-65), with leader election and no central server. |\n🪟 An app platform |\nA Paw can ship its own UI, rendered as a panel under the dashboard's Apps tab — the way apps live on a desktop. The control plane becomes a self-hosted app you extend, one panel per paw. |\n\nA Paw can ship its own interface: it drops a static HTML file in its package, declares it in the\nmanifest, and the control plane renders it as a sandboxed panel under the **Apps** tab — one entry\nper paw that has a UI.\n\nIt's the app-platform model — self-contained apps embedded inside a host shell, the way apps live on\na desktop or extensions add panels to an editor: the panel calls its paw's own tools directly (proxied over IPC — **no LLM in\nthe loop, no tokens, no extra port**), so a paw author gets a real interactive app — dashboards,\nforms, live data views — running deterministically inside the dashboard. The reference example,\n`paw-markets`\n\n, embeds a live watchlist with sparklines and alerts.\n\nThe further this goes, the less OpenVole is \"an agent runner\" and the more it's a **dashboard you\nextend into a full app** — every capability, agentic or not, a panel you can attach.\n\n**The agent loop** — the only thing the core does natively:\n\n```\nBootstrap → Perceive → Compact → Think → Act → Observe → loop\n```\n\n**Paws** — subprocess-isolated plugins that connect OpenVole to the world (APIs, databases,\nbrowsers, messaging). Each runs behind a capability-based permission sandbox.\n\n**Skills** — behavioral recipes: a folder with a `SKILL.md`\n\nof instructions, optionally with\nbundled scripts the agent runs sandboxed to the skill's own directory. Install from\n[VoleHub](https://github.com/openvole/volehub) (OpenVole's registry, with per-file hash\nverification) or [ClawHub](https://clawhub.ai) (13,000+ skills).\n\n**Brain** — the LLM lives in a Paw, not the core. `@openvole/paw-brain`\n\nis one unified paw for\nall providers and auto-detects from your API keys.\n\nDeep dive → [Architecture](https://openvole.github.io/openvole/architecture).\n\nA growing catalog, all installed from [PawHub](https://github.com/openvole/pawhub) via npm and\nsandboxed by default:\n\n**Brain**· unified multi-provider (`paw-brain`\n\n)**Channels**· Telegram, Slack, Discord, WhatsApp, MS Teams, Voice (Twilio)** Tools**· Browser, Shell, Filesystem, MCP bridge, Email/Resend, GitHub, Calendar, TTS/STT, Computer use, Database, Scraper, PDF, Image, Social (X/LinkedIn)**Infrastructure**· Memory (hybrid semantic + keyword), Session, Compact\n\n```\nvole paw add @openvole/paw-telegram\n```\n\nFull list & docs → [Paws](https://openvole.github.io/openvole/paws).\n\nConnect OpenVole instances across machines into a mesh. Remote tools appear in your local registry (the Brain can't tell the difference), brainless workers delegate thinking to a coordinator, and memory/sessions sync — all authenticated with hybrid post-quantum signatures (Ed25519 + ML-DSA-65), with leader election and load balancing. Eight topologies, one protocol, no central server.\n\nAgents talk to each other, too: one Brain can message a peer agent (`net_message`\n\n), or you can\nchat with a connected peer directly from the dashboard's VoleNet tab. Every remote action — tool\ncalls, brain delegation, chat — requires a signed message from an authorized peer; tools and brain\nare never exposed to peers unless you explicitly grant them.\n\n```\nvole net init my-instance        # generate an Ed25519 identity\nvole net show-key                # share your public key\nvole net trust \"vole-ed25519 ...\" # trust a peer\n```\n\nArchitecture patterns & setup → [VoleNet docs](https://openvole.github.io/openvole/volenet).\n\nSandboxed by default. Every Paw runs as a subprocess under Node's permission model — network,\nfilesystem, and child-process access are denied unless you grant them, and effective permissions\nare the *intersection* of what a Paw requests and what you approve. Optional Docker isolation, and\nan AES-256 encrypted vault for secrets.\n\nDetails → [Security](https://openvole.github.io/openvole/security).\n\nBoth are open-source agent frameworks with a shared skill format (`SKILL.md`\n\n), heartbeat pattern,\nand MCP ecosystem — skills written for one work on the other. The difference is philosophy:\n\n| OpenVole | OpenClaw | |\n|---|---|---|\nCore |\nMicrokernel — empty, ~60KB, everything is a plugin | Batteries-included — 25 built-in tools, ~8MB |\nBrain / LLM |\nExternal Paw; core is LLM-ignorant | Configurable provider in core |\nMemory |\nSource-isolated (user / paw / heartbeat scoped) | Shared |\nIsolation |\nNode permission sandbox on by default + capabilities | Optional Docker sandbox |\nNetworking |\nVoleNet P2P mesh across machines | Single machine |\nServer |\n`vole serve` control plane for a fleet of agents |\nGateway web UI |\n\nIf you like the microkernel approach — every piece a Paw you can swap, extend, or build yourself — come try it, build a Paw, write a Skill, and help this little vole grow.\n\nFull reference at ** openvole.github.io/openvole**:\n\n[Getting Started](https://openvole.github.io/openvole/getting-started)·\n\n[Configuration](https://openvole.github.io/openvole/configuration)·\n\n[CLI](https://openvole.github.io/openvole/cli)·\n\n[Dashboard](https://openvole.github.io/openvole/dashboard)·\n\n[Paws](https://openvole.github.io/openvole/paws)·\n\n[VoleNet](https://openvole.github.io/openvole/volenet)\n\nContributions welcome — see [CONTRIBUTING.md](/openvole/openvole/blob/main/CONTRIBUTING.md), and the\n[PawHub guide](https://github.com/openvole/pawhub/blob/main/CONTRIBUTING.md) for building Paws.\n\nIf it connects to something, it's a Paw.If it describes behavior, it's a Skill.If the agent calls it, it's a Tool.", "url": "https://wpnews.pro/news/show-hn-openvole-4-5-is-out", "canonical_source": "https://github.com/openvole/openvole", "published_at": "2026-07-10 09:21:54+00:00", "updated_at": "2026-07-10 09:35:24.674854+00:00", "lang": "en", "topics": ["ai-agents", "ai-infrastructure", "ai-tools", "developer-tools"], "entities": ["OpenVole", "Gemini", "OpenAI", "Claude", "xAI", "Ollama", "VoleNet"], "alternates": {"html": "https://wpnews.pro/news/show-hn-openvole-4-5-is-out", "markdown": "https://wpnews.pro/news/show-hn-openvole-4-5-is-out.md", "text": "https://wpnews.pro/news/show-hn-openvole-4-5-is-out.txt", "jsonld": "https://wpnews.pro/news/show-hn-openvole-4-5-is-out.jsonld"}}