OpenClaw

About OpenClaw

MIT-licensed personal AI assistant that runs as a local Gateway on your own machine and talks to you through WhatsApp, Telegram, Slack, Discord, Signal, iMessage and other channels, with an agent loop, file-based memory, installable skills, and support for local model backends including Ollama.

Best for: People who want one always-on assistant they can message from the apps they already use, running on their own hardware, with the freedom to point it at a local model or a cloud provider.

Deployment: Install with the project's script (curl -fsSL https://openclaw.ai/install.sh | bash on macOS, Linux or WSL2; a PowerShell installer for Windows) or with npm (npm install -g openclaw@latest, Node 22.22.3+, 24.15+, or 25.9+), then run openclaw onboard --install-daemon. The onboarding wizard sets up a model provider and can register the Gateway as a background service (LaunchAgent on macOS, systemd user service on Linux/WSL2, Scheduled Task on Windows). openclaw gateway status checks the daemon and openclaw dashboard opens the local Control UI. Telegram and WebChat ship with the core install; other channels are official plugins.

Skill level: Intermediate

Tradeoffs

This is an agent with shell, file and messaging reach, and the security docs assume one trusted operator per Gateway with no multi-tenant isolation; anyone who can message the bot can ask it to act, so lock down who may reach it before you widen its tools. The Ollama docs note that some local models handle simple prompts but struggle with the full agent tool surface, and warn that using the OpenAI-compatible /v1 URL instead of the native endpoint breaks tool calling. Memory search defaults to OpenAI embeddings unless you configure a local embedding option. The install script and npm package pull in Node, so it is heavier than a single binary. The project moves quickly, with a beta channel alongside stable releases.

Related guides and resources

Explore step-by-step setup guides, comparisons, and stack recipes for this tool category.

Best for

People who want one always-on assistant they can message from the apps they already use, running on their own hardware, with the freedom to point it at a local model or a cloud provider.

Why use it

OpenClaw is an agent you run rather than a library you build with. Its own README describes it as a personal AI assistant that runs on your devices and meets you in the channels you already use, connecting models, tools, messaging channels and optional companion apps through one Gateway. That Gateway is the local control plane: it binds to loopback by default, keeps the agent's memory as plain Markdown files in a workspace on your disk, and loads skills as SKILL.md instruction files. The provider layer is broad, and local backends are documented as first-class: the feature list names custom and self-hosted provider support for vLLM, SGLang, Ollama, llama.cpp, LM Studio, and any OpenAI-compatible or Anthropic-compatible endpoint, with a dedicated Ollama setup page. Meta's Muse Glimmer 30B model card also names OpenClaw as one of the agentic orchestration patterns the model works across (vendor-reported, August 2026), which makes it a natural place to try that model as a local agent brain.

Key features

• One local Gateway process owns your channels, sessions, tools and events; the CLI, terminal UI and web Control UI all connect to it, and it listens on loopback by default
• Messaging-first: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage and more, so the assistant lives where you already talk
• Local and self-hosted model backends are documented, not bolted on: Ollama (native /api/chat, default http://127.0.0.1:11434), vLLM, SGLang, llama.cpp, LM Studio, and any OpenAI-compatible or Anthropic-compatible endpoint, alongside hosted providers
• Memory is plain Markdown in ~/.openclaw/workspace by default, with optional hybrid vector plus keyword search when an embedding provider is configured; the docs are explicit that there is no hidden state
• Skills are SKILL.md instruction files following the AgentSkills spec, installable from the ClawHub registry, plus browser automation, shell exec with sandboxing, cron scheduling and webhooks

Common AI use cases

• Run a private assistant on a home server and talk to it from Telegram or WhatsApp
• Give a local Ollama model real tools: browser, shell, files, cron jobs, and web search
• Schedule recurring tasks such as morning digests or inbox and calendar routines through the built-in cron and heartbeat scheduling

Who should use it

• Self-hosters who want an assistant that lives in their chat apps rather than a browser tab
• Local AI users who want to point an agent at Ollama, LM Studio, vLLM or llama.cpp instead of a cloud key
• Tinkerers comfortable with a terminal, a background service, and reviewing which tools an agent may call

Who should not use it

• Anyone unwilling to grant an agent shell, file and messaging access on their machine
• Teams needing shared, multi-user isolation on one instance; the project documents one trusted operator per Gateway
• People who want a plain local chat window; a desktop client such as Jan or LM Studio is simpler

Tradeoffs

This is an agent with shell, file and messaging reach, and the security docs assume one trusted operator per Gateway with no multi-tenant isolation; anyone who can message the bot can ask it to act, so lock down who may reach it before you widen its tools. The Ollama docs note that some local models handle simple prompts but struggle with the full agent tool surface, and warn that using the OpenAI-compatible /v1 URL instead of the native endpoint breaks tool calling. Memory search defaults to OpenAI embeddings unless you configure a local embedding option. The install script and npm package pull in Node, so it is heavier than a single binary. The project moves quickly, with a beta channel alongside stable releases.

Alternatives

• nanobot
• Gemini CLI