What is MCP, and how does your AI talk to BrewInTH through it?

MCP is everywhere in AI tooling right now — in Claude Code, in Cursor, in tool after tool. If you have seen the word and wondered what it is, whether you need to install something, and what it has to do with putting a website online, this one is for you.

Here is MCP in plain language: how an AI uses it to "talk" to BrewInTH, what it can do there, how safe that is, and when you should set it up yourself (short answer: usually you don't need to).

In short
✅ MCP is a common language that lets an AI call another service's "tools" — think of a standard power socket
✅ BrewInTH offers 25 MCP tools: status, uploads, releases, logs, databases, domains and the portal
✅ The AI uses the same project token as the ready-made message, limited to that project and the permissions you gave
✅ Usually there is nothing to configure: paste the ready-made message and the AI picks MCP or the REST API itself
✅ For a permanent link, an OAuth-capable client can connect through BrewInTH's approval page

What is MCP?

MCP stands for Model Context Protocol, an open standard for how an AI and an outside service should talk to each other.

Picture power sockets. If every house had a different shape, every appliance would need its own adapter. MCP agrees on one shape of plug: any service that offers an MCP "socket" can be used by any AI that speaks MCP.

What a service exposes over MCP are called tools. Each one has a name, a description and the shape of the data it takes. The AI reads the list and decides on its own which tool to call, and when.

Two ways for an AI to work with BrewInTH

BrewInTH gives an AI two ways in, with the same token:

• The REST API: ordinary HTTP requests. An AI that can run commands calls it with curl or a small script, with the token in the Authorization: Bearer header.
• MCP, at BrewInTH's /mcp address. An MCP-capable AI sees every tool with its description and calls them like functions.

Both reach the same back end, with exactly the same permission and quota rules; only the way of talking differs. The guide behind the link you hand your AI describes both, so the AI uses whichever its tooling handles best.

Which MCP tools does BrewInTH offer?

There are 25 today, in six groups:

• Status and site checks: project_get_context reads the whole project state, site_check fetches a page of the live site from BrewInTH's side, project_insights shows visitors and errors.
• Uploads and releases: upload_create, upload_get_status, file_write_text, deployment_plan (the pre-release check), deployment_apply and operation_get (follow a queued job).
• Logs and commands: logs_tail, job_run (run a command in the project's sandbox), schedule_create and schedule_list (scheduled jobs).
• Databases: database_create, database_connection, database_query.
• App settings and secrets: secret_list, secret_set, secret_set_many, secret_delete.
• Domains, mail and the portal: domain_add, domain_verify, mailbox_create, portal_get, portal_update.

The sequence an AI uses most is check → upload → plan → release → follow → open the site: a complete deployment with nobody pressing a button.

How safe is it?

Letting an AI touch a live site sounds scary, but you stay in control at every layer.

• One token, one project. A token for project A cannot touch project B, nor your account, billing or anything else.
• 12 permissions to choose from, such as reading status, writing files, releasing, running commands, databases, domains, mail and secrets. The automatic button grants all of them; to narrow them, set the token up yourself on the AI connections page.
• Tokens expire. A token from Create the AI link and token (automatic) lasts 24 hours; your own can last from 1 hour to 7 days.
• Revoke any time. Press Revoke on the AI connections page; jobs that AI left queued are cancelled too.
• Secret values need their own permission. Reading them requires secrets.read.

Some things cannot be done with a token at all: naming the site, buying a plan or saving a Git repository token happen only in the owner's console.

Do I need to set up MCP myself?

Usually not. The normal way is:

1. Open the project and press Create the AI link and token (automatic).
2. Copy the ready-made message into your coding agent.
3. The AI reads the guide behind the link and chooses MCP or the REST API by itself.

That works with every agent that can run commands, with no configuration files to edit.

When a permanent MCP connection makes sense

If you use the same client with the same project every day and would rather not issue a token each time, a client that supports MCP with OAuth can connect to BrewInTH directly. Roughly:

1. Add an HTTP MCP server in your client, pointing at https://www.brewinth.com/mcp.
2. The client opens a BrewInTH page where you sign in.
3. Pick the project under Grant access only to, tick the permissions you want, and press Allow access.

After that the client holds a token that renews itself, good for about 7 days after its last use, and you can revoke it on the AI connections page like any other. How to add an MCP server differs from client to client; check its documentation or ask your AI.

If your project runs on a server other than the main one, its /mcp address is different — use the one the guide behind your project's link gives.

Prompts that make good use of MCP

Once the AI is connected to the project, try requests like these.

A health check:

Use the BrewInTH tools to check this project's site:
read the project status, fetch the home page and /api/health with site_check,
look at the last 7 days of statistics for 5xx errors, and list what needs fixing.

Finding the cause of an outage:

The site answers 500. Read the latest app logs on BrewInTH and find the cause,
fix the code, run deployment_plan, then release and open the site to confirm it is fixed.

Many settings at once:

Load the settings from this project's .env.example into BrewInTH secrets in one go.
Ask me for any value you don't know, never put secrets in the code,
then release again so the app reads them.

Watch the video (in Thai)

▶ From creating a project to opening the site (1:30)
Watch on YouTube

FAQ

Do I need to install anything?

No. BrewInTH's MCP is online at /mcp; there is nothing to install on your computer.

Can ChatGPT or Claude in a chat app use this MCP?

A complete deployment uploads files from your code folder, so it suits coding agents on a computer such as Claude Code, Codex, Cursor or Windsurf. Without one, use Upload and go live to send a ZIP or paste the code from your chat instead.

MCP or the REST API: which is better?

The result is the same. MCP-capable AIs tend to prefer MCP because they see every tool with its description; an AI that can only run commands does just as well with the REST API.

The token expired mid-task. Do I start over?

No. Issue a new token and paste the short message from the Continue earlier work (the old token expired) box into the same chat; the AI carries on straight away.

Can the AI delete my project or buy things for me?

No. The MCP tools have no command to delete a project, buy a plan or change your account; those happen only in the owner's console.

Where can I see what the AI did?

In the project's release history and activity in the console. Every action an AI takes is recorded.

Summary

MCP is the standard plug that lets an AI use BrewInTH's tools the way you would press the buttons yourself — checking the site, releasing, reading logs, handling the database — but only inside the project and permissions you granted. For most people the ready-made message is all it takes; daily users can link a client permanently through OAuth.

Start your first project free →

Related articles

• Set up Claude Code, Codex, Cursor or Windsurf for BrewInTH
• Deploy failed or site slow? Let AI read the logs and fix it