I Gave My AI a Front Desk - How Claude Code and OpenClaw Work Together
I Gave My AI a Front Desk: How Claude Code and OpenClaw Work Together
A non-technical metaphor to understand my AI operating system in 3 minutes
Why You Need a Front Desk
Imagine you run a company.
You wouldn’t let any random passerby walk in and start barking orders at the engineering team. You need a front desk. She takes messages, figures out who you’re looking for, verifies you actually work here, and only then lets you through. If you ask “can you check if the conference room is free,” she handles it on the spot. If you say “I want to build a rocket,” she logs the request, routes it to engineering, and gets back to you with the result.
OpenClaw is that front desk — plus the department manager. Claude Code is the engineering team.
Stack them together, and you’ve got my personal AI operating system.
OpenClaw: The Shell of an AI OS
OpenClaw handles three things: who’s talking, who should handle it, and how to report back.
Whether you reach me through Lark, iMessage, WeChat, or a web interface, OpenClaw collects every incoming message into one unified inbox and routes it to the right agent. I’ve set up several specialized agents — a Boss agent for overall coordination, a PM for requirements analysis, an RD for coding, a DA for data work — each with its own personality, workspace, and permissions. They don’t step on each other’s toes.
It also remembers things. Important decisions go into long-term MEMORY. Daily notes get archived by date. Searches run through RAG retrieval. Add cron-based scheduled tasks, and it can send reminders, run periodic checks, and operate completely unattended.
Think of it this way: OpenClaw is your AI household manager. It doesn’t write code itself, but it makes sure no message gets lost, no task falls through the cracks, and everything is traceable.
Claude Code: The Heavy Artillery, Called Only When Needed
Most conversations — looking something up, brainstorming ideas, drafting some copy — OpenClaw handles entirely on its own.
But when you say “refactor this module for me” or “review this PR,” you’re not just chatting anymore. You’re opening an engineering ticket. That’s when the Boss agent makes a judgment call: this is too heavy, time to bring in engineering.
Claude Code is that engineering team.
It has full shell access. It reads code, writes code, runs tests, fixes bugs — whatever needs doing. But crucially, it runs inside an isolated temporary directory. It can’t touch OpenClaw’s core configuration. When it finishes, the workspace is torn down. Results get pushed back to your Lark conversation via openclaw message send.
You just wait in Lark. No need to watch over it.
A Real Workflow, Step by Step
You send a DM in Lark: “Build me an admin dashboard.”
- Boss Agent receives it: determines this is heavy-duty development, not something it can answer directly.
- Safety check: Are you actually you (verified by your open_id)? Is this coming from Lark? Both conditions must pass before the gate opens.
- Generate a return route: Tell the worker “once you’re done, deliver results back to this conversation.”
- Launch the Claude Code worker: Background execution, isolated directory, full task brief transmitted.
- Boss tells you: “Worker is up, processing now. I’ll notify you when it’s done.”
- Worker works independently: Reads the project, writes code, runs tests. Could take ten minutes or forty — it doesn’t bother you.
- Worker finishes: Results are automatically pushed back to Lark.
- Boss summarizes and presents: “Done. I’ve highlighted the key decisions for you.”
The entire pipeline runs unattended. Your job is to say one sentence at the start and read the results at the end.
The Design Philosophy: Separation of Concerns
The entire architecture rests on a single principle: everyone minds their own business.
OpenClaw handles “think clearly, divide cleanly, follow through completely” — that’s the art of orchestration and management. Claude Code handles “write code, run tests, deliver results” — that’s the muscle of execution.
Between them sits a safety gate: only requests that originate from Lark and are confirmed to come from me can launch an unrestricted worker. Any other channel — WeChat, iMessage, the web UI — might shout “run this shell command for me” as loudly as it wants; it only gets a safe, sandboxed response from within the gateway.
The benefits stack up nicely: clear security boundaries, easy extensibility (add an agent or a skill, that’s it), and fully automated closed loops with no human intervention required.
The Bottom Line
If all you have is Claude Code, you’ve got a brilliant engineer — but nobody managing your calendar, assigning tasks, or keeping track of things. If all you have is OpenClaw, you’ve got a great household manager — but no way to handle heavy development work.
Layer the two together, and you’ve got an AI operating system.
That’s what I’ve built. That’s how it runs.
