Practical guides
Claude Code on a server: an AI agent that works around the clock
Installed on a laptop, an AI agent stops as soon as the lid closes. On a virtual private server (VPS), it runs continuously, executes scheduled tasks and belongs to the company. Here is how to set it up, and with which guardrails.
3 min read
Updated The SynAI team
Key points
- On a laptop, the agent stops with the computer. On a VPS, it runs 24/7.
- You install Claude Code, a versioned working repository, scheduled tasks and connectors to your tools.
- Before anything else: protected secrets, human approval on sensitive actions, a log of everything the agent does.
Why isn't an agent on a laptop enough?
An agent installed on a workstation inherits its limits: it stops with the computer, only one person can use it, and it leaves with them.
| On a laptop | On a server (VPS) | |
|---|---|---|
| Availability | When the computer is on | 24/7 |
| Scheduled tasks | Not reliably | Every morning, every Monday, every month |
| Access | One person | The whole team, from their own tools |
| Continuity | Leaves with the employee | Belongs to the company |
Availability
- On a laptop
- When the computer is on
- On a server (VPS)
- 24/7
Scheduled tasks
- On a laptop
- Not reliably
- On a server (VPS)
- Every morning, every Monday, every month
Access
- On a laptop
- One person
- On a server (VPS)
- The whole team, from their own tools
Continuity
- On a laptop
- Leaves with the employee
- On a server (VPS)
- Belongs to the company
What do you install on the server?
- Claude Code, with an API key owned by the company and a monitored usage budget.
- A versioned working repository: the agent's instructions, its document templates, its scripts and the history of what it produces.
- Scheduled tasks that start the agent at set times, with a precise instruction and deliverable.
- Connectors to your tools: email, CRM, accounting, shared files.
Which tasks should you give it?
- The morning summary. At 6 AM, the agent reads the night's emails and CRM activity, and sends you the list of priorities.
- Recurring reports. Every Monday, a two-page note on your sector, your competitors and your sales, always in the same format.
- The request queue. The team drops requests in a channel or a spreadsheet; the agent handles them and returns the result with a record of what it did.
Which guardrails should you put in place?
- API keys in a secrets manager, never in the code.
- Human approval before any sensitive action: emailing a client, changing an invoice, deleting anything.
- A dated log of everything the agent reads, writes and sends.
- Access limited to what is strictly needed.
- A maintained server: key-based access, updates, backups.
How much does it cost?
Two items. The server: a few dollars a month at most hosting providers, a fixed amount. API usage: pay as you go, based on the volume of text processed. The first month gives you a real figure, measured on your own volume.
Where do you start?
- Pick one recurring task that always follows the same logic.
- Run it on the server for two or three weeks.
- Measure the quality of the output, the time saved and the real cost. If all three hold, add a second task.
Frequently asked questions
- Do you need technical skills to use it?
- To install it, yes. To use it, no: the team drops requests and gets the results in its usual tools.
- Who can see what the agent does?
- The server is in your name, access is limited to what is needed, and every action is written to the log.
