Deploy and Hosting MCP Servers Compared (2026): Vercel, Cloudflare, Railway, Cloud Run, Firebase, AWS and Netlify

• Vercel: deploy, build and runtime logs, environment variables, domains and DNS records. Hosted at https://mcp.vercel.com with OAuth. Platform free tier: Hobby, $0/month, for personal, non-commercial use.
• Cloudflare: the whole Cloudflare API (including DNS and Workers) through search and execute tools, plus a docs tool. Hosted at https://mcp.cloudflare.com/mcp with OAuth or an API token. Platform free tier: Workers Free, 100,000 requests per day.
• Railway: list and create projects, redeploy, and a railway-agent tool for log analysis and service configuration. Hosted at mcp.railway.com, or local with railway mcp local. Auth: your railway login, or OAuth. Platform free tier: a 30-day trial with $5 in credits, then the Free plan with $1 of free credit per month.
• Google Cloud Run: list and inspect services and deploy from an image, an archive or file contents. Hosted at https://run.googleapis.com/mcp with OAuth and IAM; there is also a local npx server that can read service logs. Platform free tier: 2 million requests a month plus CPU and memory allowances (request-based billing).
• Firebase: firebase_deploy, App Hosting and Cloud Functions logs, and many other Firebase tools. Local only (npx firebase-tools mcp), using your Firebase CLI login. Platform free tier: the Spark plan, with Hosting at 10 GB storage and 360 MB/day transfer.
• AWS: any AWS API your IAM credentials allow, through aws___run_script, plus AWS docs search. Hosted at regional endpoints such as https://aws-mcp.us-east-1.api.aws/mcp, with OAuth or SigV4. Platform free tier: $100 in credits for new accounts, with up to $100 more.
• Netlify: create, manage and deploy projects, and manage environment variables and secrets. Hosted at https://netlify-mcp.netlify.app/mcp, or local with @netlify/mcp. Platform free tier: $0 forever, with a 300 credit limit per month.
• Prices and free tiers were checked on each vendor's pricing or docs page on 2026-10-23 and change often.

Vercel MCP is a remote server with OAuth. It is available on all Vercel plans. Vercel says only reviewed and approved AI clients can connect; Claude Code, Cursor, Codex CLI, VS Code with Copilot and Windsurf are on the list. After you add it, you still need to sign in and grant access to each team you want the agent to use.

claude mcp add --transport http vercel https://mcp.vercel.com
# then type /mcp inside Claude Code to sign in

{
  "mcpServers": {
    "vercel": {
      "url": "https://mcp.vercel.com"
    }
  }
}

• Deploy: deploy_to_vercel (small projects from inline files, limit 300 files), create_deployment (from a Git source or files), cancel_deployment.
• Logs: get_deployment_build_logs (with an errorsOnly option), get_runtime_logs and get_runtime_errors.
• Environment variables: filter_project_envs, create_project_env, edit_project_env, get_project_env.
• Domains and DNS: get_records, create_record, update_record, replace_domain_dns_records, get_domain_config.
• Safety: Vercel says connecting grants the AI the same access as your Vercel user account, and recommends human confirmation in your workflows. There is no read-only switch in the setup docs; limit which teams you grant during OAuth.
• Gotcha: get_project_env returns the decrypted value of a variable, so secrets can end up in the chat. deploy_to_vercel only takes file contents; Vercel says to prefer the Vercel CLI for local projects.

• Vercel MCP server: install for every agent

Cloudflare now points everyone to one server, the Cloudflare API MCP server, which it calls the only MCP server you need to work with Cloudflare. It uses Code Mode: the agent writes JavaScript that searches the API spec and then calls the API, so the whole API costs about 1,000 tokens of context. The older product servers, including Workers Bindings, Workers Builds, Observability and DNS Analytics, are marked deprecated.

{
  "mcpServers": {
    "cloudflare-api": {
      "type": "http",
      "url": "https://mcp.cloudflare.com/mcp"
    }
  }
}

/plugin marketplace add cloudflare/skills

• Tools: three in total. search finds endpoints in the API spec, execute calls the API with cloudflare.request(), and docs searches Cloudflare's developer docs.
• Auth: OAuth by default, where you pick the permissions to grant. For CI, send a Cloudflare API token as a Bearer token. Account tokens need the Account Resources : Read permission. Tokens with Client IP Address Filtering are not supported.
• Safety: there is no read-only mode in the docs. The server can make changes, so grant only the permissions you need at OAuth time, or use a scoped API token.
• Option: add ?codemode=false to the URL for one tool per endpoint (about 2,500 tools), at about 244k tokens of context. The docs tool stays available.
• Other hosted servers: Documentation (https://docs.mcp.cloudflare.com/mcp), Container (https://containers.mcp.cloudflare.com/mcp) and Browser Run (https://browser.mcp.cloudflare.com/mcp).

• Cloudflare MCP server: install for every agent

Try Railway free →Partner link

Railway's server runs at mcp.railway.com. The railway mcp command connects to it over stdio and reuses your railway login, so Claude Code needs the Railway CLI (version 5.44.0 or later). Editors that support OAuth can connect to https://mcp.railway.com directly. Project tokens are not accepted.

claude mcp add railway railway mcp
# or let the CLI write the config:
railway mcp install --agent claude-code

• Remote tools: whoami, list-projects, create-project, list-services, redeploy, accept-deploy, railway-agent and feature flag tools.
• railway-agent handles tasks like log analysis, debugging and service configuration. Railway's example prompts also pull environment variables into a .env file and assign a domain.
• Local server: railway mcp local runs without reaching mcp.railway.com and has a different tool set, including destructive tools such as remove_service, delete_domain and remove_volume. It returns a preview before it needs confirm: true.
• Install flags: --agent (pick one tool), --local, --oauth (direct https://mcp.railway.com endpoint) and --remote (already the default).
• Safety: destructive actions are marked at the protocol level, so clients that respect the hints ask for confirmation. Railway lists no read-only flag and suggests limiting use to non-critical environments where possible.

• Railway MCP server: install for every agent

Google runs a remote Cloud Run MCP server at https://run.googleapis.com/mcp. It is turned on when you enable the Cloud Run Admin API. It uses OAuth 2.0 with IAM and does not accept API keys. To add it to Claude as a custom connector you create your own OAuth client ID and secret in the Google Cloud console. There is also a local server on GitHub that runs with npx and your gcloud login.

"mcpServers": {
  "cloud-run": {
    "command": "npx",
    "args": ["-y", "@google-cloud/cloud-run-mcp"]
  }
}

gcloud auth login
gcloud auth application-default login

• Remote tools: get_service, list_services, deploy_service_from_image, deploy_service_from_archive (up to 250 MiB) and deploy_service_from_file_contents (up to 50 MiB).
• Local tools: deploy-file-contents, list-services, get-service and get-service-log (logs and error messages), plus deploy-local-folder, list-projects and create-project when run locally.
• Local env vars: GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_REGION, DEFAULT_SERVICE_NAME, SKIP_IAM_CHECK.
• Safety: the run.readonly OAuth scope (https://www.googleapis.com/auth/run.readonly) only allows reading data; the run scope allows read and write. IAM allow and deny policies can control MCP tool use, and the remote server needs roles such as roles/mcp.toolUser and roles/run.developer.
• Gotcha: SKIP_IAM_CHECK is true by default, which makes deployed services public. The remote server has no log tool; Google runs a separate Cloud Logging MCP server at https://logging.googleapis.com/mcp.

• Google Cloud Run MCP server: install for every agent

The Firebase MCP server is part of the Firebase CLI and runs locally over stdio. It uses the same credentials as the Firebase CLI: a logged-in user or Application Default Credentials. There is no hosted endpoint for the whole server.

claude mcp add firebase npx -- -y firebase-tools@latest mcp

{
  "mcpServers": {
    "firebase": {
      "command": "npx",
      "args": ["-y", "firebase-tools@latest", "mcp"]
    }
  }
}

• Deploy: firebase_deploy deploys the services in firebase.json; firebase_deploy_status checks a background deploy; firebase_init sets up services.
• Logs: apphosting_fetch_logs (App Hosting service or build logs) and functions_get_logs (Cloud Functions logs from Cloud Logging).
• Flags: --dir sets the folder that holds firebase.json, and --only turns on only the feature groups you list (core tools always load).
• Safety: there is no read-only mode in the docs. Use --only to load fewer tools.
• Gotcha: the docs list no tool for environment variables or secrets, and none for DNS.

• Firebase MCP server: install for every agent

The AWS MCP Server is a managed remote server. It runs Python scripts with AWS API access using your AWS credentials, so it can do anything your IAM role allows: deploy, read CloudWatch logs, change config or DNS. AWS says the Agent Toolkit for AWS has no additional charge; you pay for the AWS resources it uses.

claude mcp add aws-mcp https://aws-mcp.us-east-1.api.aws/mcp --transport http

claude mcp add-json aws-mcp '{"type":"stdio","command":"uvx","args":["mcp-proxy-for-aws-cli@latest","https://aws-mcp.us-east-1.api.aws/mcp","--metadata","AWS_REGION=us-west-2"],"env":{}}'

• Main tools: aws___run_script (Python in a sandbox with AWS API access), aws___get_tasks, aws___get_presigned_url, aws___search_documentation, aws___read_documentation, aws___retrieve_skill, aws___list_regions and aws___get_regional_availability.
• Auth: OAuth needs the AWSMCPSignInOAuthAccessPolicy on your role. SigV4 needs AWS CLI 2.32.0 or later, aws login, and uv (the setup runs uvx mcp-proxy-for-aws-cli).
• Region: the endpoint region picks the server; AWS_REGION in --metadata sets the default region for operations, otherwise us-east-1.
• Safety: AWS points to SigV4 if you need read-only mode that hides write tools. The MCP Proxy for AWS has a --read-only flag (off by default) that disables tools which may need write permissions. IAM condition keys can also restrict agent actions.
• Gotcha: use a role with narrow permissions. Multi-profile switching works only with SigV4.

• AWS MCP server: install for every agent

Netlify documents an official MCP server on docs.netlify.com and recommends its remote endpoint, because it always has the newest capabilities. Its README says agents can create, manage and deploy Netlify projects and manage environment variables and secrets. The docs we checked do not list tool names.

claude mcp add --transport http netlify https://netlify-mcp.netlify.app/mcp

claude mcp add netlify -- npx -y @netlify/mcp

• Local server needs Node.js 22 or higher.
• Auth: Claude Code can prompt you to log in, or you can run netlify login. If sign-in fails, the README says you can add a Personal Access Token as NETLIFY_PERSONAL_ACCESS_TOKEN, warns not to commit it, and says to remove it once the problem is fixed.
• Safety: no read-only mode is documented.

• Add an MCP server in every AI coding agent

• Next.js or front-end apps, and the agent should also fix DNS and env vars: Vercel.
• DNS, Workers or anything else on Cloudflare: the Cloudflare API server with a scoped token.
• A full-stack app with a database and little setup: Railway.
• Containers on Google Cloud, with a read-only option: Cloud Run (remote server with the run.readonly scope).
• A Firebase app: Firebase MCP with --only to keep the tool list short.
• Already on AWS: the AWS MCP Server, with SigV4 and --read-only for review-only work.
• Static sites and simple functions on a free plan: Netlify.
• Most of these servers can change production. Turn on approval prompts in your agent, and grant the smallest set of teams, scopes or IAM permissions.
• Details were checked against vendor docs and pricing pages on 2026-10-23.

• Set up MCP servers, rules and permissions in one paste
• Add an MCP server in every AI coding agent