OpenClaw

How the Gateway fits together

The Gateway coordinates sessions, channels, tools and events. You can talk to the assistant through supported messaging channels and use optional nodes or companion apps for device capabilities. Model-provider configuration belongs to this environment; the Gateway is not itself a language model. Keeping it reachable and available is part of running the assistant.

What v2026.9.5 changes

OpenClaw v2026.9.5 adds atomic-update preflight, plugin installation without a Gateway restart, shared conversations, browser pages beside the agent, guided specialist-agent setup, GPT Live meetings and phone calls, and reliability fixes. The release also makes Astra the default when no model is selected. Treat this as shipped scope, not proof that every platform, plugin, meeting or upgrade path works in every environment.

Memory goes beyond saved chat

OpenClaw’s current memory documentation describes Markdown-backed context such as USER.md, MEMORY.md and daily notes, alongside indexed memory_search and memory_get retrieval. Dreaming can distill material into longer-lived context. It would be misleading to reduce this system to manual notes or to assume only Hermes can retain useful information between sessions.

• Use explicit preferences and factual project context as the starting point.
• Check what is retrieved on a later task, rather than treating file persistence as proof that the right memory was used.
• Review distilled material when facts change; memory mechanisms can preserve stale assumptions.

Skills and recurring work

Skills package instructions and supporting resources; OpenClaw discovers them from documented locations including workspace, personal and bundled sources. Gateway cron persists job definitions, state and history and can deliver to chat or a webhook. The Gateway must be running when work is due: a saved schedule does not execute while its host is unavailable.

A practical first setup

Use the official onboarding instructions to configure one provider and one private channel. Give the agent a bounded task, such as drafting a source-linked update from several public pages. Review the sources and output before adding skills, device access or a recurring schedule.

• Start with a small workspace containing only the material needed for the trial.
• Confirm which tools run on the Gateway host and configure sandboxing where the job needs isolation.
• Test schedule delivery, failure visibility and recovery after a restart before depending on recurring output.

What running OpenClaw costs

OpenClaw is MIT-licensed software. Total cost can include model or API usage, hosting, connected services and the time spent configuring and maintaining the Gateway. Local model use changes the bill but introduces hardware and operating costs. There is no universal cheapest setup without a workload and a required quality level.

Security defaults and fit

OpenClaw’s security guide assumes one trust boundary per Gateway, rather than hostile tenants sharing a service. It documents pairing and allowlists for channel access, and sandboxing is off by default: main tools can execute on the Gateway host. Use the documented security audit and deployment guidance before granting broader access. OpenClaw fits people who value this control; a managed product may suit people who want less infrastructure work.

Useful first jobs

• Receive and deliver personal-assistant work through a configured private messaging channel.
• Build a recurring report with persistent job history and explicit delivery rules.
• Combine reviewed skills and device capabilities in a user-controlled workspace.