OpenClaw review, pricing and limits

OpenClaw suits people who want one self-hosted agent reachable from the chat apps they already use, with the model of their choice behind it. It is free and MIT-licensed, but you run the server, and sandboxing and exec approvals are off until you configure them, so it is a trusted single-operator tool by default.

OpenClaw is an open-source personal-agent gateway under the MIT licence, with more than 390,000 GitHub stars on 4 October 2026. One self-hosted Gateway connects 50-plus chat channels to models from Anthropic, OpenAI, Google or local runtimes. The OpenClaw Foundation, a 501(c)(3), stewards it and sells no paid tier.

Maker: OpenClaw Foundation · Autonomy: semi autonomous · Maturity: BETA

Underlying models: Anthropic Claude, OpenAI GPT, Google Gemini, Local models via Ollama, vLLM and llama.cpp

About OpenClaw

OpenClaw is an open-source personal-agent gateway, not a coding-only tool. You run one long-lived Gateway process on a laptop, a Raspberry Pi or a VPS, connect it to the chat apps you already use, and it routes each message to the AI model you choose. It was created by Peter Steinberger and the community, its GitHub repository dates from November 2025, and it is stewarded by the OpenClaw Foundation, an independent 501(c)(3) that employs the core team and signs the releases. The project's own docs say the Foundation is funded by donations and has nothing to sell: no subscription, no hosted version, no token.

The Gateway is the product. Telegram and WebChat ship in the core install, and Discord, Slack, WhatsApp, Signal, iMessage, Microsoft Teams, Matrix and more than a dozen other channels install as official plugins with one command. Behind the channels, OpenClaw talks to model providers such as Anthropic, OpenAI and Google, or to local runtimes such as Ollama, vLLM and llama.cpp. It can also hand a turn to a vendor's own agent loop: one plugin drives Codex, another the GitHub Copilot SDK, a third an installed Claude Code executable, while OpenClaw owns the channels, sessions and policy. If you want that kind of coding agent on its own, compare it with Claude Code and OpenAI Codex.

Extension happens through plugins and ClawHub, the project's skill registry. OpenClaw is an MCP client and server, speaks the Agent Client Protocol for editors and the A2A protocol for other agents, and can expose an OpenAI-compatible HTTP API. Companion apps cover a macOS menu bar app, a Windows Hub installer, and iOS and Android nodes for camera, voice and device commands. Setup needs Node 24.16+ or 26.1+ and a one-line installer script on macOS, Linux or Windows.

Cost is the model bill, nothing else. The software is MIT-licensed with no enterprise edition, so the only money you spend is on the model providers you connect (for example Anthropic or OpenAI API usage) and whatever hosting you choose. Running it against a local model keeps that near zero but needs capable hardware.

Where it falls short matters for this choice. By the project's own account, isolation and command approvals are opt-in, so a default install is a trusted single-operator assistant that holds your credentials and can run commands; hardening is deliberate configuration. One Gateway is one trust domain, native plugins run in-process without a sandbox, and there is no hosted option for people who do not want to run a server. A closed cloud agent such as Manus or a managed alternative like Hermes Agent is the better fit if you want zero setup or a vendor to carry the operations. Pick OpenClaw when you want control over where the agent runs and which models and channels it touches.

Pricing

OpenClaw itself is free: the software is MIT-licensed and the project's docs state the OpenClaw Foundation offers no paid plan, hosted version or token. You pay the model providers you connect by their own usage rates, plus any server you run it on. Third parties such as [Hermes Agent](/hub/agents/hermes-agent) sell managed subscriptions instead.

Key Features

• One Gateway, many channels: A single self-hosted Gateway connects Telegram, WebChat, Discord, Slack, WhatsApp, Signal, iMessage, Teams, Matrix and more through core and plugin channels.
• Bring any model: Routes to Anthropic, OpenAI, Google and other providers, or to Ollama, vLLM, llama.cpp, LM Studio and any OpenAI-compatible endpoint you host.
• Vendor agent loops as runtimes: Codex, the Copilot SDK and Claude Code can run underneath a chat session, with the Gateway still deciding who may talk to what.
• ClawHub skills and plugins: A skill registry with per-release security scans, plus plugins that add channels, memory, voice and tools through documented SDK contracts.
• Policy as code: Tool availability and exec denial are enforced in the Gateway rather than requested in the prompt, with secret handles that keep protected credentials out of model context.
• Companion apps and nodes: Menu bar, Windows installer and phone nodes add pairing, voice, camera and device commands.

Strengths

• MIT-licensed with no enterprise edition, so every feature is in the free download.
• Model-agnostic: swap Anthropic, OpenAI, Google or a local model without changing the channels or workflows around it.
• Meets you inside the chat apps you already use instead of a new interface.
• Releases are signed and published under the Foundation identity, with a public advisory list and a written trust model.

Weaknesses

• A fresh install holds your credentials and can run commands until you switch on isolation and approvals.
• You operate the server, the updates and the model bills; there is no hosted tier to offload that to.
• One Gateway is one trust domain, and native plugins run in-process without a sandbox.
• Fast release cadence (date-versioned releases roughly weekly) means pinning a release channel for stable deployments.

Visit OpenClaw Official Site