Deutsch
OpenClaw is an open-source AI assistant that connects a model to messaging channels, tools and persistent workspace context. You can use it to receive requests, gather information, draft responses or carry out configured tasks from an environment you operate.
For example, an assistant could receive a request in a chat channel, read an approved project source and return a briefing. Sending messages, editing files or changing another system requires additional permissions. Start with the task you need, then give the assistant the access needed to do it.
The OpenClaw self-hosting guide covers installation on a remote server.
What OpenClaw provides
OpenClaw brings together an assistant runtime, model connections, tools, skills and ways to communicate. The upstream repository describes a Gateway that manages sessions, tools, events and channel connections.
The Gateway is the service around the assistant. Interfaces such as its control UI, command line and messaging channels connect to it. Your configuration determines which model it uses, what context is available and which tools it can call.
Self-hosting lets you choose where this runtime and its state live. Hosted-model requests and connected messaging services still send data to the providers you configure, so review those data flows separately.
How a request becomes a task
A typical workflow has a few steps:
- An authorised user sends a request through a configured channel or interface.
- The runtime supplies the relevant conversation and workspace context.
- The model selects an answer or a tool action.
- The configured permissions determine whether that action can proceed.
- The assistant uses the result and reports back.
Stored context can help across sessions, but it can also be stale. A reminder about last month's deployment does not establish the state of the server today. Ask for current evidence when the answer depends on a changing system.
If the Gateway or a dependency is unavailable, the workflow may stop. Plan restart behaviour, logs and error reporting alongside the assistant's instructions.
Messaging, tools and skills
Messaging integrations let requests arrive through familiar channels. Configure who may issue commands and how unfamiliar senders are handled. A connected group chat should not automatically grant every participant access to private files or tools.
Tools connect the assistant to actions such as reading files or calling an API. Skills provide reusable instructions. Review both the skill content and the access of the tools it invokes; a helpful description does not establish that a third-party integration is trustworthy.
The upstream security guidance notes that tools can run on the host unless sandboxing is configured. Confirm the effective setup before connecting other users or sensitive resources. Use the OpenClaw security documentation for the release you run.
Useful starting workflows
Briefings and meeting preparation
Combine approved project notes, calendars or status sources into a dated briefing. Include references and distinguish missing information from a confirmed absence of activity. Review what private context may appear in the destination channel.
Inbox and message triage
Summarise conversations, suggest priorities and draft replies. Keep sending as a separate action that requires approval until the audience, account and task rules are clear. A draft is easy to review; an incorrectly sent message is harder to recover.
Developer assistance
Gather issue context, inspect a repository or prepare a change in a dedicated workspace. Use normal code review, tests and acceptance checks. Keep production credentials and merge or deployment decisions outside an exploratory task.
Routine business reports
Read a queue, collect approved metrics or prepare draft content. Define what counts as a completed report and how failures should be surfaced. Avoid letting a repeated run create duplicate records or notifications.
Home-lab experiments
Connect a small set of tools to disposable data and observe the results. Test denied access as well as successful actions before adding more systems. This provides a useful way to learn the runtime without making a complex workflow the first experiment.
How it differs from a basic chat interface
OpenClaw gives you a configurable runtime around model interaction, including channels, tools and stored state. Other assistants may offer overlapping features, so compare the workflow and control you need rather than assuming that every chat product lacks memory or tools.
The main operational difference appears when the assistant can act. A wrong answer needs correction; a wrong tool call can alter a file, expose information or update another service. Permissions and review should reflect that difference.
When a VPS helps
A VPS can keep the Gateway reachable when a laptop sleeps, disconnects or changes networks. It also provides a dedicated place for configuration, logs and stored context, separated from personal desktop files.
That separation depends on the setup. Use a dedicated account, restrict SSH and administrative interfaces, protect credentials and limit network access. Configure the runtime to recover from a restart and monitor failed requests, disk growth and resource use.
A server still needs updates and independent backups. Model usage and connected services can add charges beyond the VPS price. Test the actual tools, because browsers, builds or local models can need much more capacity than a simple chat workflow.
Use a local workspace when it meets the need more simply. A remote server is worthwhile when persistent availability or access from several devices justifies the extra operation.
Prepare access and recovery
Before enabling a serious workflow, write down the allowed users, channels, directories and tools. Assign credential owners and approval rules for sending, deletion, billing or changes to shared systems.
Keep useful logs without storing unnecessary secrets or private content. Define retention, backup and restore steps for the workspace state. Know how to stop the runtime and revoke an integration if a task behaves unexpectedly.
On a self-managed VPS, your team operates OpenClaw and the host. Introduce access gradually, and verify results before turning a manual task into a recurring one.
Frequently asked questions
What is OpenClaw useful for?
It suits assistant workflows that combine messaging, context and tools: briefings, drafts, research and bounded recurring tasks. Suitability depends on the integrations and review process you configure.
Must it run on my own computer?
No. A compatible remote server can host the runtime. Check upstream platform requirements and any tools that need a local device or desktop integration.
Does self-hosting keep all data local?
No. The runtime can store state on your server while sending prompts or tool data to configured model providers, chat services and integrations. Review each connection.
Can it run without supervision?
Low-risk recurring tasks can be configured with clear limits and failure reporting. Retain approval for consequential actions, and do not treat unattended execution as proof that the results are correct.
