---
title: "Hosting for AI coding agents in 2026: let Claude Code or Codex deploy and debug production | Pethost"
description: "To let Claude Code or Codex deploy and debug production alone, a host must give the agent five things. Pethost does, with 13 MCP tools, from €9.99 a month."
---

[Blog](https://pethost.dev/blog/) · Hosting for AI agents

# Hosting for AI coding agents in 2026: let Claude Code or Codex deploy and debug production

By the Pethost team · 16 September 2026 · Updated 5 October 2026 · 10 min read

You built an app with an AI coding agent and just want it online, with Claude Code or Codex deploying it to production by itself and finding out why it broke. That takes a host that gives the agent five things: a sign-in with no token to copy, logs and requests it can afford to read, a rollback, limits on what it may do, and a price that stays put. We built Pethost around those five: an MCP server with 13 tools on a machine of your own, for €9.99 a month.

**In short**

- **Three ways to let an agent deploy:** a CLI with a token, a push to GitHub with CI, or the host's MCP server. The third is the one made for a model, if its tools reach past the deploy: logs, requests, a rollback.
- **Most large platforms have an official MCP server in October 2026**, Railway, Render, Vercel, Netlify and Cloudflare among them. They differ in what an agent can do alone: Railway's hosted server hands log analysis to Railway's own agent, Render's and Netlify's have no rollback, Vercel's has 213 tools.
- **Pethost's server has 13 tools:** deploys, every request for 30 days, logs filtered and paged, a rollback in one step, backups. The agent signs in through your browser and cannot pay or change your account.
- **The price does not move when the agent deploys ten things:** €9.99 a month is a machine of 2 vCPU, 3 GB and 30 GB. The same 3 GB of always-on memory is $30.00 a month at Railway's rates and $39.00 at Render's plan prices.

## What an AI coding agent needs from a hosting platform to deploy and debug production

An agent cannot click through a dashboard, pays in tokens for every line it reads, and can do whatever its credentials allow. Five needs follow.

### A sign-in without a token to copy

A token pasted into a terminal lives on in the shell's history, a config file and sometimes the chat. With OAuth the agent opens your browser and you approve it once. On Pethost that is the only way in, and you revoke it in Settings at any time.

### Logs and requests in one place, in answers a model can afford

To debug production an agent asks which path fails, and with what: the request log and the container log, from one server. It pays for every answer and for the list of tools: Cloudflare counts a tool for each endpoint of its API at “~244k tokens vs ~1k tokens” for the two it ships instead. On Pethost `QueryHttpTraffic` answers which paths fail and which are slow, `QueryContainerLogs` what a request failed with. Logs come filtered and paged, and one call waits for a deploy.

### A rollback

An agent that deploys in one sentence will, some evening, deploy a broken version in one sentence. The way back has to be a tool too. On Pethost `DeployProject` takes an earlier deploy or an earlier commit, and the volumes keep their data.

### Limits on what it may do

The agent should run your projects, not your account. Vercel's documentation says its server “grants the AI system you're using the same access as your Vercel user account”; Railway's advises: “Avoid production risks by limiting usage to non-critical environments where possible.” On Pethost the agent can deploy, read logs and requests, change settings, restart, roll back and restore. It cannot pay for anything or change your account.

### A price that does not move when an agent deploys ten things

When a deploy costs one sentence, you deploy more: the bot, the staging copy, Saturday's idea. Where every service has its own price or meter, each is a new line on the bill. On Pethost the machine is the price.

## Three ways to let Claude Code deploy your project to production by itself

**Three ways an agent deploys**

| Way | What the agent holds | What it learns after the deploy |
| --- | --- | --- |
| A CLI with a token | An API token in a file or a variable | What the CLI prints for a terminal |
| A push to GitHub with CI | The Git access it already has | Nothing, unless a second tool reads the host |
| An MCP server | A sign-in you approve and can revoke | What its tools return: the deploy's result, logs, requests |

A CLI works with any agent that has a shell, and some MCP servers are the CLI again: Fly.io's is a command of flyctl marked “[experimental]”. A push needs nothing new, but the agent does not see the build fail, or the 500s after it, until you tell it. On Pethost every push to GitHub deploys itself too, and the agent asks the same server what happened.

## What if I just give the agent SSH to a VPS?

You can. Two things first. An agent on a bare server works as root over SSH: nothing tells it what runs there, and one wrong command takes down the proxy, a database or the firewall. On Pethost its shell is inside your containers, never on the machine.

And a server is not sure to cost less. Hetzner, the one most guides name, raised its prices twice in 2026, and its cheap Cost-Optimized range has long been out of stock. Those servers are hardware it bought at good prices, before the RAM shortage, and on 5 October every plan of the range was marked “not available”. The 2 vCPU, 4 GB server it does sell, CPX22, is €19.49 a month, up from €7.99, before a panel, backup storage and your hours.

## Official MCP servers in 2026: Railway, Render, Vercel, Netlify and Cloudflare

**What an agent can do through each official MCP server, October 2026**

| Host | Deploys your code | Logs | Requests | Rollback | Restores a backup |
| --- | --- | --- | --- | --- | --- |
| Pethost | A folder or a GitHub repository | Filtered and paged | Every request, 30 days | An earlier deploy or commit | Any kept night |
| Railway | Templates and staged changes | Through `railway-agent`, or the local server | On the local server | None named | None named |
| Render | From a Git repository | App and build logs | With a Pro workspace, $25 a month | None | None |
| Vercel | A Git source or files | Runtime logs, build output | No tool of their own | `request_rollback` | None |
| Netlify | `deploy-site` | None | None | None | None |
| Cloudflare | No deploy tool: code against the API | Workers' logs, on two more servers | No tool of their own | None named | None named |

All six can sign the agent in through your browser; Vercel takes only clients it has approved. More hosts are in [hosting providers with an official MCP server](https://pethost.dev/blog/hosting-with-mcp-server/).

### Railway and Render

Railway's hosted server has 11 tools, and one, `railway-agent`, passes “a natural-language request to Railway's AI agent for multi-step operations like log analysis, debugging, and service configuration”: your agent asks another agent. Render's server has 26: one, `list_logs`, reads app, build and request logs, and three return a link to the dashboard.

### Vercel, Netlify and Cloudflare

Vercel's server has 213 tools. Netlify's deploys a site, and the checklist in its source leaves “rollback/deploy selection” unticked. Both are made for front ends and functions: a Vercel function ends after 300 seconds by default and a Netlify background function after 15 minutes, so a bot or a worker runs somewhere else. Cloudflare's main server reaches its whole API, “over 2,500 endpoints”, through two tools: the agent writes JavaScript against the API where another server has a tool named deploy.

Self-hosted panels have servers too: Coolify's has 45 tools, most of them reading, and Dokploy's mirrors its API in “508 tools across 49 categories”. The server under the panel stays yours to update and back up: [the self-hosted panels compared](https://pethost.dev/blog/coolify-vs-dokploy-vs-caprover-vs-dokku/).

## How Claude Code or Codex deploys and debugs on Pethost, step by step

### 1. Install the plugin

Subscribe, and a machine is set up for you. The plugin is a skill that teaches the agent to deploy and look after projects, and the MCP server's address.

Claude Code

**You say:** Install the Pethost plugin from github.com/pethost-dev/plugin for yourself.

**The agent answers:** Installed. Approve me in your browser: there is no token to copy.

- GetMachine 2 cores, 3 GB memory, 30 GB disk, no projects yet

**The agent answers:** I see your machine. What shall I deploy?

### 2. Deploy

The agent packs the project, writes a Dockerfile if there is none, and deploys it as a Docker Compose project, which we run [as written](https://pethost.dev/blog/managed-docker-compose-hosting/).

Claude Code

**You say:** Deploy this and send me the link.

- CreateTransfer website.tgz uploaded
- CreateProject website: building
- GetOperation succeeded

**The agent answers:** It's live at https://website.sam.pethost.app

### 3. Find why a form fails

Two calls: the requests name the path, the logs name the error.

Claude Code

**You say:** Why does the contact form fail?

- QueryHttpTraffic POST /api/contact: 7 requests, all 500
- QueryContainerLogs Error: SMTP_HOST is not set

**The agent answers:** The app wants SMTP_HOST, and the project's .env has none. Which server?

**You say:** smtp.example.com. Set it.

- DeployProject /.env changed: succeeded

**The agent answers:** Set and deployed. Try the form.

**What the agent read, you can read.** The same requests in the panel: the busiest paths, and each request under them.

![A project's requests in the Pethost panel: its ten busiest paths, each with its share, its count and its response times, and the log of single requests under them.](https://pethost.dev/shots/traffic.webp)

### 4. Roll back

The machine keeps the files of earlier deploys, and one call puts them back. A GitHub project goes back to an earlier commit the same way.

Claude Code

**You say:** The new footer broke the page. Roll back to yesterday's version.

- GetProject current deploy: today 14:02; files kept: yesterday 18:40
- DeployProject rollback to yesterday 18:40: succeeded

**The agent answers:** Back on yesterday's version. Your .env and the volumes stayed as they are.

## What €9.99 a month buys when an agent deploys ten things

Ten small things an agent puts online need memory that stays on, not ten plans. Starter has 3 GB of it. The same 3 GB, always on, at each host's own plan prices:

**3 GB of always-on memory, per month. Bars compare at €1 = $1.1204, the ECB's reference rate of 5 October 2026**

| Host | Price a month |
| --- | --- |
| Pethost Starter | €9.99 |
| Fly.io | $20.10 |
| Railway | $30.00 |
| Render | $39.00 |

Our €9.99 is the best value here, the whole machine: 2 vCPU, 30 GB of disk, traffic, backups and logs are inside it. Railway's $30.00 is memory alone, at $10 a gigabyte: CPU, disk and traffic come on top. Render's $39.00 is one instance of 2 GB and two of 512 MB. Fly.io's $20.10 is three Machines of 1 GB in Ashburn, before volumes and a database.

Counted by projects, three always-on web services, a worker and a Postgres are $34.75 a month on Render and, by memory alone on our assumption of use, $23.15 on Railway. They fill about 2.5 GB of Starter's 3 GB, and the next project costs nothing more while it fits: [unlimited projects for one flat fee](https://pethost.dev/blog/unlimited-projects-flat-monthly-fee/), with [no usage-based billing](https://pethost.dev/blog/paas-without-usage-based-billing/).

- **2 vCPU, 3 GB, 30 GB** a machine of your own
- **13** MCP tools for your agent
- **30 days** of every request
- **14 nights** of backups, off the machine

Also done for €9.99, without VAT: HTTPS on every address, a deploy on each push to GitHub, SSH into any container, security updates. There is no free tier: the price rents a real machine, so nothing sleeps. [Grow](https://pethost.dev/#pricing) is twice the machine for €17.99.

## Who Pethost is for

### Pethost is the best pick for

- **People who build with Claude Code, Codex or Cursor** and want the agent to deploy and run what it wrote
- **Solo developers with several side projects**: one bill of €9.99, however many the agent deploys
- **Early startups** that iterate fast and want one fixed bill
- **Developers tired of looking after a server**

### Look elsewhere if

- Your application needs several regions or autoscaling: here one machine is the limit

## Questions

### How do I let Claude Code deploy my project to production by itself?

Give it a host with an MCP server and sign it in once. On Pethost you tell Claude Code to install the plugin from github.com/pethost-dev/plugin, approve it in your browser and say “deploy this and send me the link”: it packs the project and answers with the address.

### Is there a hosting service made to work with AI coding agents like Claude Code or Codex?

Yes. Pethost was designed for agents from the first day: an MCP server with 13 tools and a skill, through which Claude Code, Codex, Cursor and others deploy, read logs and requests, roll back and restore, on a machine of your own for €9.99 a month.

### I built an app with an AI coding agent and just want it online. Where should I deploy it?

Where the same agent can also look after it. On Pethost that is one sentence to the agent and €9.99 a month for anything Docker runs, with HTTPS, nightly backups and request logs included. There is no free tier.

### Do Railway, Render, Vercel, Netlify and Cloudflare have an official MCP server in 2026?

Yes, all five had one on 5 October 2026. They differ in what an agent can do alone: Railway's hosted server passes log analysis to Railway's own agent, Render's and Netlify's have no rollback, and Vercel's has 213 tools. Pethost's has 13, with logs, requests and a rollback among them.

### Is it safe to let an AI agent deploy to production?

As safe as the limits the host puts on the agent: a sign-in you can revoke, no way to pay or change the account, and a rollback one tool away. On Pethost all three hold.

Read from each vendor's own documentation and pricing page on 5 October 2026. Tool counts are ours. Prices are in each vendor's currency; ours are without VAT.

1. [Hetzner Cloud: Cost-Optimized plans](https://www.hetzner.com/cloud/cost-optimized)
2. [Hetzner: price adjustment](https://docs.hetzner.com/general/infrastructure-and-availability/price-adjustment/)
3. [Railway: MCP server](https://docs.railway.com/ai/mcp-server)
4. [Railway: pricing](https://railway.com/pricing)
5. [Render: MCP server](https://render.com/docs/mcp-server)
6. [Render: its repository](https://github.com/render-oss/render-mcp-server)
7. [Render: pricing](https://render.com/pricing)
8. [Vercel: Vercel MCP](https://vercel.com/docs/agent-resources/vercel-mcp)
9. [Vercel: its tools](https://vercel.com/docs/agent-resources/vercel-mcp/tools)
10. [Vercel: function limits](https://vercel.com/docs/functions/limitations)
11. [Netlify: agent setup](https://docs.netlify.com/build/build-with-ai/agent-setup-guides/agent-setup-overview/)
12. [Netlify: its repository](https://github.com/netlify/netlify-mcp)
13. [Netlify: pricing](https://www.netlify.com/pricing/)
14. [Netlify: background functions](https://docs.netlify.com/build/functions/background-functions/)
15. [Cloudflare: its MCP servers](https://developers.cloudflare.com/agents/model-context-protocol/cloudflare/servers-for-cloudflare/)
16. [Cloudflare: its repository](https://github.com/cloudflare/mcp)
17. [Cloudflare Workers: pricing](https://developers.cloudflare.com/workers/platform/pricing/)
18. [Fly.io: fly mcp server](https://docs.fly.io/flyctl/cmd/fly_mcp_server)
19. [Fly.io: pricing](https://docs.fly.io/about/pricing)
20. [Coolify: MCP server](https://coolify.io/docs/mcp/capabilities/overview)
21. [Dokploy: its repository](https://github.com/Dokploy/mcp)

## Put your projects online today

A machine of your own, as many projects as fit, one flat price. Nothing metered, nothing billed on top.

[Host all your projects for €9.99/mo](https://console.pethost.dev/)

## Keep reading

- 1 October 2026 [Always-on hosting for $10 a month in 2026: no cold starts, no sleeping apps](https://pethost.dev/blog/always-on-hosting-no-cold-starts/) About $10 a month keeps one 512 MB service always on at Render or Heroku. On Pethost €9.99 rents a 3 GB machine where nothing sleeps: room for six.
- 27 September 2026 [European hosting for an app with Postgres at a flat monthly price in 2026: what €10 a month buys](https://pethost.dev/blog/european-hosting-postgres-flat-price/) An app with Postgres in the EU at a flat monthly price is €9.99 on Pethost: a 3 GB machine for every project that fits. Eleven hosts, priced.
- 21 September 2026 [Hosting with unlimited projects for one flat monthly fee in 2026: a price per machine, not per app](https://pethost.dev/blog/unlimited-projects-flat-monthly-fee/) Hosting priced per machine runs as many projects as fit for one flat monthly fee. Six small apps: €9.99 on Pethost, $30 on Railway, $42 on Render.
- 17 September 2026 [PaaS with no usage-based billing in 2026: flat prices, spending caps and overage compared](https://pethost.dev/blog/paas-without-usage-based-billing/) Of six PaaS that bill by use or per service, only Railway documents a hard spending cap. Pethost has no usage-based billing: €9.99 a month for a whole machine.

[All articles](https://pethost.dev/blog/) · [About Pethost](https://pethost.dev/about/) · [Pricing](https://pethost.dev/#pricing)

Pethost · Stephan Gomer · NIP 1133096755

[Home](https://pethost.dev/) [About](https://pethost.dev/about/) [Blog](https://pethost.dev/blog/) [Terms](https://pethost.dev/terms/) [Privacy](https://pethost.dev/privacy/) [Refunds](https://pethost.dev/refunds/) [support@pethost.dev](mailto:support@pethost.dev)