ZeroClaw review

ZeroClaw is a self-hosted agent runtime distributed as a single Rust binary, not a hosted API or managed SaaS. It sits between model providers, messaging channels and tools: messages from configured channels reach one agent loop, which can call tools such as shell, browser, HTTP, hardware interfaces and custom MCP servers. The README names Anthropic, OpenAI, Ollama and about 20 other providers, including OpenAI-compatible endpoints.

You install it on a machine you control, configure it, and can run it interactively or as a background service. The CLI offers quickstart, agent and service commands; the README lists systemd, launchctl and Windows Service registration. ZeroClaw also describes a web dashboard for chat, memory browsing, configuration, cron management and tool inspection. Its scope extends to event-triggered procedures using MQTT, webhooks, cron or peripherals. This makes it an orchestration layer for a personal or project agent, rather than a model or channel provider.

The primary integration surface is the runtime and its CLI, not a named client SDK. The README documents zeroclaw quickstart, zeroclaw agent -a, and service install and start commands. Configuration connects providers, channels and tools; the runtime can route inbound messages to one agent loop. For extensions, it names custom MCP servers, HTTP tools, webhooks and MQTT-triggered procedures. The homepage does not name a public REST API or SDK languages; confirm SDK support and any API contract in the vendor's docs before designing an application around one.

The installation documentation covers Unix, Windows, source builds and platform-specific paths, and the project links a full quick-start walkthrough and architecture material. The README says the Unix installer can be inspected with ./install.sh --help. Rust is the implementation language, and the project ships a single Rust binary; that does not establish which client languages are supported. Reddit feedback on setup is mixed: commenters describe successful script installs as well as permission, configuration and documentation difficulties. Validate your exact channel and provider path in the docs before rollout.

ZeroClaw is presented as a self-hosted, open-source runtime, and the pricing page lists a free plan. It lists no paid plans, starting price, monthly or annual billing terms, add-ons or overage rates. That makes the runtime’s published price straightforward, but it does not establish the total cost of running an agent.

You supply the machine or other compute environment, and your configuration connects to model providers such as Anthropic, OpenAI or Ollama. The ZeroClaw pricing page does not publish model-provider charges, request allowances, rate limits, seat charges or usage-based billing. Local Ollama is a listed provider option, but the project does not publish a cost comparison for local and hosted inference. Budget separately for infrastructure and any provider charges, and confirm those providers’ current terms for your workload. Before production, ask whether any paid support or service terms exist; the pricing page lists none. There is no published ZeroClaw price curve to use for estimating a bill as message volume grows.

The repository includes Apache-2.0 and MIT licence files, and the source is available on GitHub. ZeroClaw is designed to run on your machine; its README says it uses your keys and workspace. That gives you control over the runtime environment rather than requiring a vendor-hosted agent service.

The README names several model providers, channels and tools, which gives you options to change parts of the stack. The project also supports custom MCP servers. Those are useful boundaries against depending on a single provider, but they do not guarantee that every configuration, memory store or workflow can move unchanged. The project does not publish a clear data-export or migration guarantee in the homepage text. Before adopting it, confirm where memory and configuration are stored, how to back them up, and which formats can be exported. You can inspect and modify the source, but custom changes and provider-specific behavior can still create maintenance work.

ZeroClaw lists security controls including supervised autonomy, workspace boundaries, command policy and OS-level sandboxes. The README says medium-risk operations require approval and high-risk operations are blocked by default. Treat those controls as configuration and runtime features to validate, not a substitute for reviewing tool permissions, secrets handling, network access and operating-system isolation. Hardware access and shell tools can expand the consequences of a misconfiguration.

The project homepage does not publish rate limits, scaling limits, uptime commitments or compliance certifications. Test provider failures, channel reconnects, backup and restore, and service restarts against your own deployment requirements. Reddit reports vary: one commenter said the install script worked with minor service and UI setbacks; other users described broken setup steps, permission errors and configuration problems. These are individual reports, not a reliability measure, but they make a controlled pilot sensible. Confirm the repository identity before installing; a Reddit user warned that similarly named impostor repositories exist.