In 2026 a store’s hosting should play four roles at once: serve the website, be operable by an AI agent (an MCP server with permission scopes), run agents (for example Hermes in its own container, working through the night too) and give them data safely (read-only databases, vectors in PostgreSQL, mailing lists). Aderlo Cloud brings these four layers together in one panel.
From a server for a website to a workplace for agents
Until recently, hosting was judged by one question: does the site run fast and reliably? AI agents add three more: can an agent safely do something on this hosting, where does the agent work when you close your laptop, and what data can it reach without putting that data at risk?
| Layer | Question | How it looks on Aderlo Cloud |
|---|---|---|
| 1. Website | Is the store fast and stable? | LiteSpeed Enterprise, LSCache, Redis per account, dedicated PHP workers |
| 2. Agent access | Can an agent work safely on the account? | MCP server with OAuth, 5 scopes, trash instead of deletion — details |
| 3. Agent environment | Where does the agent work 24/7? | The Hermes Workspace app in its own container |
| 4. Data for the agent | What can the agent use? | PostgreSQL with pgvector and an agent-access switch, MySQL databases, NetSendo |
What an agent can realistically do in a store today
- Maintenance: plugin updates through WP-CLI, checking mail records, database clean-ups, tracking down errors in files.
- Sales and analytics: reading orders from the database (read-only) and reporting what changed and why.
- Content and SEO: product descriptions, structured data, internal linking.
- Email marketing: welcome sequences and abandoned-cart emails in your own NetSendo.
- Automation: fixing n8n workflows, morning reports sent to Telegram.
A 30-day plan
Week 1: read-only access
Connect the agent over MCP with the “account” scope and read-only database access. Have the agent describe the state of the store: domains, resource usage, the largest tables.
Week 2: first tasks with write access
Add the files or DNS scope for one specific job (for example SPF/DMARC records). Keep confirmation on for destructive tools.
Week 3: an agent that works on its own
Launch Hermes with a role (for example sales analyst or SEO specialist) and schedule a weekly report.
Week 4: data for the agent
For search and for answering customer questions, set up a PostgreSQL database with pgvector; give the agent access to it only where it is needed.
Rules that protect the store
- One agent = one connection = one scope. Easy to revoke, easy to see who did what.
- The agent reads production data in read-only mode; write access only where the agent builds something of its own.
- A backup before every larger piece of agent work, even if nightly backups exist.
- An agent with broad access should not process untrusted content (customer emails, web pages) unsupervised — it may contain planted instructions.
Visibility: agents choose providers too
More and more buying decisions start with a question put to an AI assistant. Hosting that an agent can operate — with a public MCP server, documented tools and clear documentation — is hosting an agent can check and recommend based on facts rather than advertising. Aderlo Cloud publishes its server in the official MCP Registry as cloud.aderlo/hosting.
Frequently asked questions
- Do I need a VPS to work with AI agents?
- No. An agent can manage ordinary hosting through an MCP server, and if it needs to work on its own around the clock, on Aderlo Cloud you run it as the Hermes Workspace app in its own container, with no server to administer.
- Which AI model should the agent use?
- Hermes Agent works with many model providers; on Aderlo Cloud the agent uses your own account with the provider, so you choose the model and the costs yourself.