Gemini built the whole thing.
Now ask for the address.
One command adds agenthost to Gemini CLI. After that, the session that started with an empty folder can end with a link that works on someone else’s phone.
A finished project that is still only on your laptop.
Gemini will happily read a whole project and rebuild half of it in an afternoon. That leaves you with something genuinely finished and nowhere to put it, and the deploy is suddenly the slowest part of the day: an account somewhere, a wizard, a build command you have to go and look up.
agenthost is the step that is missing. Connect it once and Gemini gets tools it can call: create a project, send the files, get back an https link. Nothing has to be configured, which is the whole trick. There is no build command field, so there is no build command for either of you to get wrong.
It reads the files and works out the rest. An index.html is a website and costs nothing. A package.json is a Node app. A requirements.txt is Python, served through the entrypoint we expect. Gemini reads the deploy reference before the first one, so it knows the conventions without you explaining them.
You ask for it in a sentence.
There are no commands to learn on our side. These are whole sentences, typed the way you would say them to a person, and each one is a thing agenthost does.
put the site online
Sends the files, works out the runtime, and answers with a working https link. Websites are the free part.
the python one too, it needs a database
Deploys the app, creates a database and attaches it, in one go.
set the API key as an environment variable
Stores it as a secret on the app. It never has to appear in your project files.
check it's actually up
Reads the deploy status and the running app’s logs, and says what went wrong if anything did.
take the old demo down
Removes it properly. What you stopped using stops costing anything.
There is more of this in the how-to guides, which are written for your agent to read as much as for you.
Whatever it built, it can publish.
You do not have to know what it is made of. Your agent works that out and tells us.
A website or page
The one-pager, the portfolio, the event invite. Files, served fast, with nothing to build first.
Ten free, forever, no card.
How it works →An app that saves things
Sign-ups, notes, bookings, orders. Somewhere to run and a database to remember, together.
€4 a month, database included.
How it works →Node, Python, PHP
Express, Flask, Django, a script that has to run. One fixed convention each, and no build command to get right.
The contract, in full.
What runs here →- framework preset
- build command
- output directory
- install command
- API keys
- credit card
- region
Nothing here can stop and ask you a question.
Every one of these is a setting some other host would demand before it would take your files. None of them exists here, which is why your agent can go from a folder to a working link without checking in.
- Dockerfiles
- port settings
- TLS certificates
- DNS records
- CI pipelines
- health checks
- scaling rules
The one thing agenthost does expect is an entrypoint per runtime, and the deploy reference is where Gemini reads it before the first deploy.
How it connects.
One command in your terminal. -s user makes agenthost available in every project rather than only the one you are standing in.
gemini mcp add -s user -t http agenthost https://agenthost.eu/mcp
Add the server
It writes agenthost into your Gemini CLI settings. There is no key or header to pass, because agenthost signs you in rather than checking a secret.
Sign in with your email
Run /mcp auth agenthost inside Gemini, or gemini mcp auth agenthost from your shell. Your browser opens, you type the code we email you, and you approve it.
Ask for a link
Say put this online. If you are new to agenthost, that same first deploy is what creates your account.
Run /mcp list afterwards to see agenthost and its tools, which is the quickest way to confirm it took. The get started page walks through all of it, with the buttons and the sign-in step.
The usual ones.
Can I use this in the Gemini web app rather than the CLI?
You can, through custom apps for Spark: on gemini.google.com, go to Settings and help, then Connected Apps, then add a custom app with the agenthost address. Google gates that fairly tightly at the moment, to Spark users on a personal Google account, and it works inside Spark tasks rather than everywhere in Gemini. The CLI has none of those conditions.
Is there an API key to configure?
No, and you should leave the header and token options alone. agenthost registers itself with Gemini automatically and then signs you in through your browser, so there is nothing to paste and nothing to rotate later.
Does it work with Gemini in an IDE?
If the surface you are using reads your Gemini CLI settings, agenthost is there too, because that is where the server is written. Anything else that speaks MCP can be pointed at the same address by hand.
What can it actually do once connected?
Create projects, deploy, read logs, roll back, attach a database, set environment variables, add a domain, invite people to something private, and open a support ticket if it gets stuck.
Is there an API key to paste anywhere?
No. agenthost signs you in through your browser, and your agent never sees a password or a key. There is nothing to copy out of one window and into another, and nothing to rotate later.
What does it cost to try?
Nothing, and there is no card. Your first ten websites are free and stay free. You only pay when the thing you built has to run code on our side, or when you want it on your own domain.
Can I connect more than one agent to the same account?
Yes, as many as you like. They all see the same projects, so you can build something in one and update it from another. Connecting the second one is a single click, because your browser is already signed in.
Where is this actually hosted?
In the EU, by a European company. Your files, your database and your logs stay here, and so does the billing.
Not what you use?
agenthost is a standard remote MCP server, so anything that speaks MCP can drive it. These are the ones with a page of their own.
Claude Code
Deploy the project you are already standing in, without leaving the terminal.
Read more →Claude
The Claude apps and claude.ai. Claude builds you something, then gives you the link to it.
Read more →Claude Cowork
Cowork works away on the thing for an hour, then publishes it at a real address.
Read more →ChatGPT
Turn on developer mode, paste one address, and ChatGPT can put things online.
Read more →Codex
The Codex CLI, app and IDE extension, all from the one config file they share.
Read more →Microsoft 365 Copilot Cowork
A plugin your admin uploads once, with the deploy skill built in.
Read more →Cursor
One button, and Cursor can publish what it just built for you.
Read more →VS Code
GitHub Copilot's agent mode, with agenthost added as an MCP server.
Read more →Point Gemini at it.
One command to connect, no card, and the project that was only on your laptop gets an address.