OpenClaw, properly deployed.
Setup, integration, and ongoing maintenance for self-hosted AI agent fleets — running on your infrastructure, under your control, with your data staying where you put it.
Running one agent is easy. Running a fleet is not.
A single bot on a laptop takes an afternoon. What takes longer is everything after: routing between model providers so an outage does not take you offline, wiring agents into the channels your team actually uses, keeping credentials out of config files and chat logs, and surviving the upgrade that silently changes a default you depended on.
Most of the difficulty in a self-hosted agent platform is not the agent. It is the operations around it.
What this covers
Setup & deployment
Gateway installation on Linux under Docker or natively on Windows. Channel integration for Microsoft Teams, WhatsApp, and others. Model routing across providers with sensible fallbacks. Autostart, logging, and a config that survives a reboot.
Integration
MCP servers connecting agents to the systems you already run — Microsoft 365 and Graph, accounting platforms, internal APIs. Custom tools and skills. Persistent memory with embeddings so an agent remembers across sessions.
Maintenance & reliability
Version upgrades validated with per-bot smoke tests rather than hope. Scripted health checks. Backups, warm-standby failover, and credential rotation before a token quietly expires and takes an integration down with it.
Runs where you need it
Linux & Docker
Containerised deployments with per-bot isolation, outbound-only ingress, and no inbound ports opened on your network.
Native Windows
Full gateway deployments on Windows without WSL, running as scheduled tasks or services, for estates that are Windows-first.
Cloud & hybrid
Cloud-hosted agents with secure tunnels, or on-premises agents with a cloud warm standby for failover.
How the work is done
Changes are validated before they are applied, not after. Configuration is version-controlled so any change can be traced and reverted. Secrets live in environment variables and credential stores, never in tracked files or chat transcripts. Where a control claims to restrict something, it gets tested by actually trying the thing it is supposed to block.
That last one matters more than it sounds. A setting that validates cleanly and silently does nothing is worse than no setting at all, because you stop checking.
Common questions
What is OpenClaw?
A self-hosted gateway for running AI agents. It connects language models to chat channels, tools, and your own systems, and it runs on hardware you control rather than someone else’s platform.
Do I need my own servers?
Not necessarily. It runs on a small Linux box, a Windows machine, or a cloud VM. Sizing depends on how many agents you run and whether you host models locally or call hosted ones.
Which models can it use?
Hosted providers such as Anthropic and OpenAI, and local or cloud open-weight models. Most useful setups mix them — a capable model for reasoning, a cheaper one for routine scheduled jobs, and a fallback so a provider outage does not stop everything.
Where does my data go?
Wherever you decide. Agents run on your infrastructure, but any request sent to a hosted model provider reaches that provider. Which data is allowed to reach which provider is a configuration decision, and one worth making deliberately rather than by default.
Can agents work in Microsoft Teams?
Yes. Teams is a first-class channel, alongside WhatsApp and others. Agents can hold conversations, respond to mentions, and send scheduled reports into channels or direct messages.
Start a conversation
Describe what you are trying to run and what is currently in your way.