Pomeroy has an HTTP endpoint. This is still true.
That is the part worth being precise about, because the endpoint sounds like it should change the answer. The endpoint binds to loopback only:127.0.0.1, this Mac’s own address,
which no other machine can route to. It is not bound to your Wi-Fi address, it
is not bound to 0.0.0.0, and there is no port forwarded to it.
A cloud-hosted assistant’s MCP client runs on its vendor’s servers. To reach a
tool it has to make a request to a public address. Pomeroy does not have
one, and the endpoint’s binding is the reason rather than an accident of
configuration.
Assistants this rules out
- claude.ai in a browser, and Claude in the cloud
- ChatGPT on the web
- Grok
- Any assistant whose tool calls originate from a vendor’s servers
What to use instead
The desktop equivalents run on your Mac, so they can. In most cases the same account and the same models.Claude Desktop
Instead of claude.ai. Free, and the shortest setup Pomeroy has.
Codex
Instead of ChatGPT on the web. Reaches the ChatGPT desktop app too.
Perplexity for Mac
Instead of perplexity.ai in a tab.
LM Studio
Or skip the cloud entirely and run the model on your Mac.
Could you tunnel it yourself?
Technically, yes. A tunnel from a public address to127.0.0.1:13579 would
work, and nothing in Pomeroy tries to detect one or stop you.
Be clear about what it costs before you do. You would be putting a public door
in front of your mail, your messages and your calendar, guarded by one token,
on a machine that has no other inbound exposure. The endpoint’s threat model
assumes an attacker already has code running as you on this Mac, which is a
much smaller assumption than assuming the whole internet can knock.
If you do it anyway, use Approval mode rather than Token, and treat
the tunnel’s own authentication as the real gate.