Blog · Hosting for AI agents

Hosting providers with an official MCP server in 2026: deploy, logs and rollback compared

By the Pethost team · 25 September 2026 · Updated 5 October 2026 · 10 min read

You want your coding agent to deploy an app, read its logs when it breaks and roll back a bad version, and you ask which hosting providers have an official MCP server. In October 2026 almost all do: 18 of the 19 platforms we read. Most can start a deploy and read some logs, two have a rollback tool, and Pethost's 13 tools do the whole job, a restore from a backup included, for €9.99 a month.

Which hosting providers have an official MCP server in 2026

We read each vendor's documentation on 5 October 2026 and, where it lists no tools, the server's source. Remote means an address the agent connects to; local, a process on your computer.

“Through execute” means one tool runs code the agent writes against the vendor's API. AWS, Azure, Zeabur and Replit have official servers too: on AWS the agent writes Python for aws___run_script, and on Replit a deploy is a prompt to Replit's own agent.

Deploy, logs and rollback: what each official MCP server's tools do

Railway MCP server: your agent asks Railway's agent

The hosted server's railway-agent tool hands “a natural-language request to Railway's AI agent for multi-step operations like log analysis, debugging, and service configuration”. A logs tool of its own, get_logs, is in the local server, railway mcp local. Neither list names a rollback.

Render MCP server: three kinds of logs, no rollback

list_logs reads app, build and request logs; request logs need a Pro workspace, $25 a month. Three update tools return only “a link to the service's settings in the Render Dashboard”, and for a rollback Render's docs say to “use the Render Dashboard or REST API”.

Vercel MCP: deployments, logs and a rollback

get_runtime_logs returns up to 1,000 entries, and request_rollback points production traffic “to a previous production deployment by ID”. The server is for what Vercel hosts: front ends and functions that run per request. Only AI clients “reviewed and approved by Vercel” connect.

Netlify, DigitalOcean App Platform and Google Cloud Run: deploy, no logs

A checklist in the source of Netlify's server leaves “rollback/deploy selection” and “builds - why did it fail” unticked. DigitalOcean's App Platform server and Cloud Run's remote one deploy, and have no tool for logs.

Fly.io, Heroku and Koyeb MCP servers: deploys, logs and releases

fly mcp server is flyctl's commands as tools, marked “[experimental]”: none does fly launch or fly deploy, and fly-apps-releases only lists releases. Heroku's local server has deploy_to_heroku and get_app_logs. Koyeb's lists “cancel deployment” as planned.

Cloudflare and Sevalla: two tools and code the agent writes

Cloudflare's API server reaches “over 2,500 endpoints” through search() and execute(): the agent writes JavaScript that calls the API. Sevalla's does the same over an API that has a rollback and a database restore: reachable as code the agent composes, not as a tool with a description.

Coolify and Dokploy: an official MCP server on a server you run

Coolify's is built into the panel since v4.1.0 of May 2026, and most of its tools read: it redeploys and restarts an application that exists, and creates none. Dokploy's “exposes all Dokploy API endpoints as tools”, its rollback described as “POST /rollback.rollback”. Under either, the server stays yours to update and back up: see the self-hosted panels compared and Coolify Cloud vs Dokploy Cloud.

An MCP server designed for agents, or an API handed to them

An agent pays for a server twice: the tool list is sent to the model with the conversation, and every answer is read as tokens.

The longest lists hold a tool for each API route. Dokploy's README: “Some MCP clients and LLM providers can be slower or less reliable when very large tool lists are sent to the model.” The shortest give the agent the API to program: a tool for each endpoint would take Cloudflare's server “~244k tokens vs ~1k tokens”. Pethost's 13 each do one clear thing, and the API reference states their rules:

• Answers sized for a model. A response holds about 64 KiB at most; a cut list says so and pages. Logs are filtered by service and text, requests by path, method and status.
• Errors that say what to do next. NOT_FOUND lists what exists, INVALID_ARGUMENT names the field and what it takes, and a refused compose.yaml comes back as violations, each saying what to write instead.
• One call that waits for a deploy. CreateProject and DeployProject wait up to 45 seconds and answer with the project as it then is: its URL, its services' states, its problems. For a longer build GetOperation waits on.

A debugging session and a rollback through Pethost's MCP server

A shop deploys from GitHub, and one morning its checkout breaks. A coding agent finds the cause in three calls:

The first call said which path fails, the second why, the third which deploy brought it. The rollback is one call:

Had the bad version damaged data, RunProjectAction restores a nightly snapshot, after backing the project up first. Postgres here is a container in the project and in that backup: not a managed database with point-in-time recovery.

What €9.99 a month buys, the MCP server included

The MCP server comes with every plan. On Starter, €9.99 a month without VAT, what its tools read is already running:

• 2 vCPU, 3 GB, 30 GB a machine nobody shares
• 13 MCP tools, and a skill that teaches the agent
• 30 days of request logs to query
• 14 nights of backups to restore

HTTPS addresses, a deploy on every push to GitHub, SSH into any container and security updates are in the price too: the best value we found for an agent's projects. There is no free tier.

An agent creates things, so it matters what each one costs. On Render create_web_service takes a plan, and every service is a line on the bill: $7 a month for 512 MB. On Pethost the agent “cannot pay for anything or change your account”: the machine is the limit, and with no usage-based billing the next project adds nothing to the bill while it fits.

Railway's figure is memory alone, with CPU, disk and traffic on top; Fly.io's is three 1 GB Machines in Ashburn ($23.19 in Frankfurt), before volumes and a database; Sliplane's is its 4 GB server, the first that holds 3 GB; Koyeb's counts its $29 plan. On a server you rent for Coolify or Dokploy an agent also has root over SSH, where one wrong command takes down the proxy; and the 2 vCPU, 4 GB server Hetzner does sell, CPX22, is €19.49 a month: its cheap range has long been out of stock, hardware bought before the RAM shortage. Railway vs Render vs Fly.io pricing has the plans; Grow doubles the machine for €17.99.

Who Pethost is for

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

1. Railway: MCP server
2. Render: MCP server
3. Fly.io: fly mcp server
4. Heroku: remote MCP server
5. Heroku: local MCP server
6. DigitalOcean: MCP tools
7. Vercel MCP: tools
8. Netlify: MCP server
9. Cloudflare: MCP servers
10. Cloudflare: API MCP server
11. Cloud Run: MCP reference
12. AWS: MCP tools
13. Azure: MCP tools
14. Koyeb: MCP server
15. Northflank: agent skills
16. Sevalla: MCP server
17. Coolify: MCP capabilities
18. Dokploy: MCP server
19. Sliplane: MCP server
20. Zeabur: MCP server
21. Replit: MCP server
22. Railway: pricing
23. Render: pricing
24. Fly.io: pricing
25. Heroku: pricing
26. DigitalOcean: pricing
27. Koyeb: pricing
28. Northflank: pricing
29. Sevalla: pricing
30. Sliplane: pricing
31. Hetzner Cloud: Cost-Optimized plans
32. Hetzner: price adjustment
33. Coolify Cloud: pricing