AI & Domains

Connect OpenClaw or Hermes Agent to Your Domains (via MCP)

2026-08-13 · OSIR Team

Summary: OpenClaw and Hermes Agent are both full MCP clients, which means either one can operate a registrar that exposes an MCP server. This guide shows the general setup, what a domain conversation with your agent looks like in practice, and - just as important - the guardrails: why a registrar connector should refuse anonymous calls, gate every purchase behind explicit confirmation, and never be paired with unreviewed third-party skills.

Part 2 of our Agents & Domains miniseries. Part 1 mapped the landscape: Meet the AI agents that can manage your domains.

What you need

An agent that speaks MCP. This guide uses OpenClaw and Hermes Agent, the two big open-source personal agents; the same steps apply to Claude, Gemini CLI, or any other MCP client.
A registrar with an MCP server. We will use our own: the OSIR connector at https://be.osir.com/mcp/http exposes 87 tools covering search, registration, DNS, transfers, VPS, app hosting, and billing. Any registrar MCP server works the same way in principle - most just expose far fewer capabilities.
An account at the registrar. This is not optional, and below we explain why you should treat it as a feature.

Step 1: add the MCP server to your agent

Both agents take remote MCP servers in their configuration:

OpenClaw supports MCP servers as first-class tool providers - add the server URL to your OpenClaw configuration (see the MCP section of the OpenClaw docs for the current syntax; it changes as the project moves fast).
Hermes Agent has a dedicated MCP integration guide in its documentation - register the server URL and the tools appear alongside its built-in ones.
Claude users: no config files at all - Settings, then Connectors, then add custom connector with the URL above.

After connecting, ask the agent to list its new tools. You should see the domain groups: availability and search, registration and transfers, DNS, contacts, billing.

Step 2: authenticate - and why the connector refuses to work without it

Here is something we verified live while writing this post: if you connect our MCP server and immediately ask for an availability check, the tool answers with "Authentication required." Even a read-only lookup is refused until the session is authenticated.

That is deliberate, and when you are wiring up an autonomous agent it is exactly what you want. An anonymous, tool-capable endpoint is an invitation for abuse; an authenticated one means every call your agent makes is tied to your account, rate-limited, and visible in your audit trail afterwards.

Two ways to authenticate:

OAuth - in clients with a connector UI (like claude.ai), you log in once when adding the connector and the session is handled for you.
Device login - for headless agents like OpenClaw and Hermes: the agent calls the loginWithDevice tool, which returns a link and code. You open the link, approve the login in your browser, and the agent receives a session key it passes on subsequent calls. The agent never sees your password.

Step 3: talk to your domains

Once connected and authenticated, domain management becomes conversation. Real workflows this setup handles:

The availability check, anywhere:

You (WhatsApp): is coffeecart.dev free? what about .com and .al?

Agent: coffeecart.dev is available. coffeecart.com is taken, coffeecart.al is available. Want pricing for the two available ones?

The DNS change without a dashboard:

You: point blog.mydomain.com at 203.0.113.7

Agent: I'll create an A record: blog.mydomain.com -> 203.0.113.7, TTL 3600. Confirm?

The expiry watchdog (this is where an always-on agent like Hermes earns its keep - it can check on a schedule, not just when you ask):

Agent (unprompted, Tuesday 09:00): Heads up: mydomain.com expires in 30 days and auto-renew is off. Turn auto-renew on, renew now, or leave it?

The registration:

You: register coffeecart.dev for 2 years

Agent: coffeecart.dev, 2 years - here is the price including the ICANN fee and payment processing. This will charge your account balance. Confirm?

Every action that spends money or is hard to reverse comes back as a staged confirmation the agent cannot skip: the connector's executeConfirmedAction gate means registration, renewal, and payment tools return a summary and wait for an explicit yes. We covered how that gate behaves in a full workflow in the deploy-everything walkthrough.

Step 4: the guardrails (read this part)

Personal agents are powerful because they hold broad permissions, and the ecosystem's security record shows what that costs: researchers have demonstrated prompt-injection attacks on autonomous agents, and malicious skills have been found in community skill registries. Connecting an agent to anything that can spend money deserves deliberate care:

Prefer registrars that authenticate every call. If a domain endpoint answers an anonymous agent, treat that as a red flag, not a convenience.
Insist on a confirmation gate for purchases. The agent proposes; you approve. If a connector can buy without a human confirming, do not give it payment capability.
Review what else the agent runs. A registrar session in the same agent as an unreviewed third-party skill is a bad combination - a malicious skill can try to steer the agent's other tools. Install skills the way you install browser extensions: sparingly and from sources you trust.
Keep the blast radius small. Use a modest account balance rather than an uncapped card, keep transfer lock and registrar lock on, and skim the audit trail occasionally - one of the connector's tools, getMyAuditLogs, lets you ask the agent to report on its own past actions.
Start read-only. Let the agent do lookups, monitoring, and reporting for a week before you let it change DNS or renew anything.

None of this is unique to domains, but domains punish sloppiness more than most things an agent touches: a wrong DNS change takes your email down, and a lapsed renewal can cost you a name permanently (see the domain lifecycle for how unforgiving the expiry pipeline is).

Why bother

Because the always-on agent changes what a registrar account is for. A dashboard you visit twice a year cannot notice anything; an agent that holds your domain tools can watch expiry, verify DNS after a migration, compare renewal pricing across your portfolio, and catch the misconfiguration before your visitors do - and it reports to you in the chat app you already have open. The setup above takes minutes; the mistakes it prevents are the expensive kind.

Setup FAQ

Does this work with agents other than OpenClaw and Hermes?

Yes. Any MCP client works: Claude (connector UI), Gemini CLI, Claude Code, Cursor, Codex CLI. OpenClaw and Hermes are featured because they are the always-on, messaging-first agents where domain automation is most useful.

Can the agent register a domain without asking me?

Not through our connector. Registration, renewal, and payment are staged behind an explicit confirmation step (executeConfirmedAction), and every call requires an authenticated session tied to your account. The agent can prepare everything; only you can approve the charge.

What happens if my agent goes haywire and spams the API?

Calls are authenticated, rate-limited, and logged per account. You can review everything the agent did via the audit-trail tools and revoke the session at any time by logging the device out.

Is there a fee to use the MCP connector?

No. The connector itself is free to use with an OSIR account - you pay only for what you buy through it (domains, VPS), at the same prices as the website, with the same fee disclosure.

Where do I find the full tool list?

The complete catalog - all 87 tools in 9 groups - is on the MCP page, and the AI integration docs cover A2A and the REST API for agents that do not speak MCP.

Part 2 of the Agents & Domains miniseries. Previous: Meet the AI agents that can manage your domains.

Image generated with AI (Higgsfield).