AI App Backend

Best MCP servers for building and deploying apps (2026)

Eight official MCP servers that let ChatGPT, Claude, Claude Code, Codex and Cursor build, deploy and run real apps, with what each does well, what to watch, and what the underlying service costs.

Floot ·

Key Takeaways

  • Floot is the only server here that covers the whole app. One MCP server gives the agent a Postgres database, sign-in, endpoints, jobs, storage, email and hosting, and lets it publish. That is why it ranks first for this reader.
  • GitHub’s server is worth adding to almost any setup. Repos, issues, pull requests and Actions, with toolsets and a read-only mode to keep the agent’s reach narrow.
  • Supabase, Vercel and Stripe are the best single-purpose picks. Database work, frontend deployments and payments respectively, each with OAuth sign-in and documented scoping or confirmation features.
  • Neon, Cloudflare and Render suit specific stacks. Neon for Postgres branching in development, Cloudflare for Workers and DNS across 2,500+ API endpoints, Render for creating services and databases.
  • Every server here acts with your account’s access. Keep tool approval on for writes, scope servers to one project where possible, and start against development data.

Quick-Start Prioritization Framework

RankServerBest forAddressUnderlying service from
1FlootBuilding and publishing a full-stack app through one serverDirectory listings; mcp.floot.com/mcpFree; Pro $25/month
2GitHubCode, issues and pull requestsapi.githubcopilot.com/mcp/Free; Team $4/user/month
3SupabasePostgres, auth and Edge Functions on Supabasemcp.supabase.com/mcpFree; Pro $25/month
4VercelFrontend deployments and logsmcp.vercel.comHobby free (non-commercial); Pro $20/month
5StripePayments, subscriptions and billing datamcp.stripe.com2.9% + 30¢ per US card payment, no monthly fee
6NeonPostgres branches for development and testingmcp.neon.tech/mcpFree; usage-based paid plans
7CloudflareWorkers, DNS and the wider Cloudflare APImcp.cloudflare.com/mcpWorkers free; paid from $5/month
8RenderCreating web services and databases on RenderSee Render docsFree instances; paid per service

An AI coding agent is only as useful as what it can reach. Connect the right MCP servers and Claude Code can open the pull request, Cursor can create the database table, and Codex can deploy and read the logs when something breaks. Connect the wrong ones, or too many, and you get slow tool selection and a wide security surface. We compared eight official, remote MCP servers that matter most when building and shipping an app, using each provider’s own documentation and pricing pages as of 2026-11-11. The list is ranked for one reader: someone who wants an agent to take an app from idea to a live URL.

How we chose and ranked

We only included official servers run by the company behind the service, because a third-party server holding your credentials is an extra party to trust. All eight are remote servers you add by URL or from a directory, with no local process to install.

  • Reach: how much of the build-and-ship path the server covers on its own.
  • Agent support: documented setup for Claude Code, Cursor, Codex and the Claude and ChatGPT apps.
  • Safety controls: OAuth sign-in, read-only modes, project scoping, confirmation steps.
  • Cost: the price of the underlying service from its own pricing page. The servers themselves don’t carry separate charges in the docs we read.

1. Floot: best for building and publishing a whole app

Floot builds and hosts full-stack web apps (React frontend, serverless endpoints, Postgres) from plain-English descriptions. Its MCP server is the main way to use it from ChatGPT, Claude, Claude Code, Codex and Cursor, and the same project can also be edited in Floot’s own editor.

What it does well

The agent creates the project, writes pages, components and endpoints, adds tables to the app’s own Postgres database, turns on sign-up and login (email, Google or Microsoft, with roles), adds scheduled jobs and background tasks, file storage, email, push notifications and realtime, checks logs, and publishes. You watch a live preview as it works. Floot Payments adds checkout and subscriptions paid out via Stripe, and Floot can build and submit native iOS and Android apps. Because one server covers all of this, there are no keys to copy between services.

What to watch

Apps are written in Floot’s structure, so an existing codebase needs adapting to move in. Code download is on paid plans; database download is on every plan, and paid plans can restore the database to any moment in the last 7 days. If you already run a stack you like on other platforms, the single-purpose servers below fit better.

How to connect

Install Floot from the Claude connector directory, the ChatGPT app directory (Codex uses the same ChatGPT account) or the Cursor Marketplace. For Claude Code or the Codex CLI, add https://mcp.floot.com/mcp manually and sign in.

Price

Free: $0, 100 building actions a day, $3 one-time hosting allowance. Pro: $25 a month, 1,000 actions a day, $10 hosting a month, 5,000 AI credits, custom domains. Power: $100 a month, 5,000 actions a day, $25 hosting, 15,000 AI credits. Publishing is free on every plan.

2. GitHub: best for code, issues and pull requests

GitHub’s official MCP server is hosted at api.githubcopilot.com/mcp/ and can also run locally through Docker. It works with any MCP host that supports remote servers.

What it does well

Tools are grouped into toolsets such as repos, issues, pull_requests, actions and code_security. The default set covers context, repos, issues, pull requests and users, and you can switch on others or individual tools. A --read-only flag skips every write tool even if one is requested. That makes it a good fit for agents that review, triage or investigate CI failures.

What to watch

With a personal access token, the token’s scopes decide the agent’s reach; GitHub suggests repo, read:packages and read:org as a minimum. Issue and pull request text written by others is untrusted input that can carry prompt injection, so be careful combining it with servers that can change production.

Price

GitHub Free is $0 with unlimited public and private repositories and 2,000 CI/CD minutes a month. Team is $4 per user a month; Enterprise is $21 per user a month.

3. Supabase: best for Postgres, auth and functions on Supabase

Supabase’s server at mcp.supabase.com/mcp gives an agent the database and platform tools for a Supabase project, and a local version runs with the Supabase CLI.

What it does well

Database tools list tables, extensions and migrations, apply migrations and run SQL. Debugging tools query logs and fetch Supabase’s security and performance advisors, which flag issues such as missing Row Level Security. The agent can also generate TypeScript types and deploy Edge Functions. Parameters limit the blast radius: project_ref scopes to one project, read_only=true allows only reads, and features picks tool groups.

What to watch

Supabase recommends connecting to production only when necessary, using branching to test changes, keeping manual approval of tool calls on, and never giving the server to your own customers. Frontend hosting is separate.

Price

Free: 500 MB database, two active projects, paused after a week of inactivity. Pro from $25 a month. Branching, which pairs well with agents, is $0.01344 per branch per hour on Pro and Team.

4. Vercel: best for frontend deployments and logs

Vercel MCP is a remote server with OAuth at mcp.vercel.com, available on all Vercel plans. Its docs list setup for Claude Code, Claude, ChatGPT, Codex CLI, Cursor, VS Code and others, and npx add-mcp https://mcp.vercel.com configures detected clients in one step.

What it does well

Search Vercel’s documentation, manage projects and deployments, analyse deployment logs and query Web Analytics. Vercel’s changelog also announces that the server can now deploy code. It is the natural companion for a Next.js frontend on Vercel paired with a separate backend.

What to watch

Vercel says connecting grants the AI the same access as your Vercel user account, and recommends human confirmation in workflows. Clients must be on Vercel’s reviewed list or identify themselves with a Client ID Metadata Document.

Price

Hobby is free for personal, non-commercial use. Pro is $20 a month, with developer seats listed at $20 a month each.

5. Stripe: best for payments and billing

Stripe’s server at mcp.stripe.com lets an agent search Stripe’s API and docs and read or write Stripe data. It is listed in the Claude connector directory and the ChatGPT plugin directory, and installs in Cursor, Claude Code, Codex and VS Code.

What it does well

Four general tools, stripe_api_search, stripe_api_details, stripe_api_read and stripe_api_write, expose a long list of API methods: customers, products, prices, payment links, Checkout Sessions, subscriptions, invoices, coupons, refunds, webhook endpoints and more. There is also a documentation search and an implementation planner. Some actions, such as creating a refund, require you to approve them in the Stripe Dashboard before they run.

What to watch

Stripe recommends turning on your client’s tool-approval setting and being cautious combining Stripe with other servers because of prompt injection. Its docs say that from December 2026 the server stops accepting full-access secret keys and restricted keys without the Agent tag; use OAuth or an agent API key. Start in a sandbox.

Price

Stripe’s standard US pricing is 2.9% + 30¢ per successful domestic card transaction, with no setup or monthly fees.

6. Neon: best for Postgres branches in development

Neon’s server at mcp.neon.tech/mcp manages serverless Postgres projects. npx neon@latest mcp writes the configuration for you, and OAuth or an API key handles sign-in.

What it does well

Create and delete projects and branches, run queries and transactions, compare schemas, and apply schema changes on a temporary branch before they reach the main one. Neon has also added tools for its auth, Data API, object storage and functions products. URL options narrow access: ?readonly=true, ?projectId=, ?category= and ?roleName=.

What to watch

Neon recommends its MCP server for development and testing only: never connect agents to production databases, use anonymised data, and approve each action. Its docs also note that roles created in the Console are members of neon_superuser, so roleName doesn’t restrict reads unless you create a narrower role.

Price

Free: 100 projects, 100 CU-hours of compute and 1 GB of storage per project. Launch and Scale are usage-based with no monthly minimum; Launch compute is $0.106 per CU-hour and storage $0.35 per GB-month.

7. Cloudflare: best for Workers, DNS and the Cloudflare API

Cloudflare’s main server at mcp.cloudflare.com/mcp covers more than 2,500 API endpoints across DNS, Workers, R2, Zero Trust and other products. Separate servers cover documentation, containers and browser rendering.

What it does well

Instead of hundreds of tools, it exposes three: search(), execute() and docs(). The model writes JavaScript against a typed description of the API, which runs in an isolated sandbox. That keeps the tool list small while giving very wide reach, which is useful if your app already runs on Workers and R2.

What to watch

Wide reach cuts both ways: an agent with this server can change DNS and security settings as well as deploy code. Use a scoped API token rather than broad OAuth access for automation, and keep approval on.

Price

Workers Free allows 100,000 requests a day. Workers Paid has a $5 monthly minimum including 10 million requests and 30 million CPU milliseconds, then $0.30 per extra million requests.

8. Render: best for creating services and databases on Render

Render’s MCP server lets Claude Code, Claude Desktop, Codex and Cursor manage Render resources with natural-language prompts, signing in through OAuth or an API key. Official Render plugins for those agents configure it automatically.

What it does well

Create web services, static sites, cron jobs, Render Postgres databases and Key Value instances; trigger deploys; update environment variables; run read-only SQL; and pull logs and metrics such as CPU, memory and response times.

What to watch

Render’s docs list what the server doesn’t cover: deleting resources, scaling settings, image-backed services and changes beyond deploys and environment variables go through the Dashboard or API. They also note that Render can’t guarantee sensitive data such as connection strings stays out of the AI app’s context.

Price

Render grants 750 free instance hours per workspace each month. Free web services spin down after 15 minutes without traffic, and free Postgres databases expire 30 days after creation. Paid instances are billed per service.

Side-by-side comparison

What each server lets an agent do, according to its docs.

ServerDatabaseAuth for your app’s usersDeploy / publishLogsRead-only or scoping option
FlootYes, Postgres per appYesYesYesTool approval in your agent
GitHubNoNoVia ActionsActions runsRead-only flag, toolsets
SupabaseYesNot listed in MCP toolsEdge FunctionsYesproject_ref, read_only
VercelNoNoYesYesOAuth grant, team permissions
StripeNoNoNoTool call logs in WorkbenchAgent API keys, confirmations
NeonYesNeon Auth toolsFunctions toolsYesreadonly, projectId, category
CloudflareVia the Cloudflare APINot listedWorkersVia the Cloudflare APIAPI tokens for automation
RenderPostgres, Key ValueNoTrigger deploysYesLimited write surface

Recommended combinations

Most people need two or three servers, not eight.

Simplest: let one platform do it

Floot alone, plus GitHub if you also keep code there. The agent builds, adds the backend and publishes without you managing keys across services.

Classic web stack

Supabase + Vercel + GitHub, adding Stripe when you start charging. The most flexible, with four accounts and four sets of permissions to keep tidy.

Edge-first stack

Cloudflare + Neon + GitHub, for apps built on Workers that want Postgres branches per feature.

Common mistakes

These are the problems that come up most often once agents have real access.

Connecting production first

Supabase, Neon and Stripe all steer you to branches, development projects or sandboxes. Point the agent there until the workflow is proven.

Mixing untrusted input with write access

A GitHub issue or a fetched web page can contain instructions. If the same session also has a server that can run SQL or refunds, keep approval on for those writes.

Using community copies of official servers

Search results and marketplaces list unofficial servers with similar names. Vercel’s docs explicitly tell you to verify the domain before connecting. Use the address from the provider’s own docs or directory listing.

Too many servers at once

Every server adds tool descriptions the model must read and choose between. Turn off servers you aren’t using for the current task.

Frequently Asked Questions

Which MCP server should I add first?

If you want the agent to build and publish a whole app, Floot. If you already have a codebase, GitHub, then the server for wherever your database and hosting live.

Do MCP servers cost money?

The official servers in this list don’t list a separate charge; you pay for the underlying service as normal, such as hosting, database usage or Stripe’s transaction fees.

Can I use these MCP servers in ChatGPT and Codex?

Yes. Floot, Stripe and Vercel document ChatGPT setup, and Codex reads remote servers from ~/.codex/config.toml or codex mcp add. Codex shares its MCP configuration with the ChatGPT desktop app.

Is it safe to give an AI agent access to my production database?

Neon recommends against it entirely for its server, and Supabase recommends production access only when necessary with read-only mode and project scoping. Use development branches and keep approval on for writes.

What is the difference between Vercel’s and Floot’s MCP servers?

Vercel’s manages projects, deployments and logs on Vercel; your database and auth come from other services. Floot’s builds the app itself, including its database, sign-in and backend, and hosts it.

Can an MCP server deploy my app?

Several can. Floot publishes apps, Vercel’s changelog says its server can deploy code, Render’s triggers deploys, and Cloudflare’s can deploy Workers through its API.

Sources