AI App Builder MCP Support (2026): Which Ones Can an Agent Actually Drive?

Quick Answer

"Does it support MCP?" is the wrong question in 2026, because MCP support is three unrelated products sharing one name, and a builder can ship any one of them without the other two.

1. Control plane. An outside agent in ChatGPT, Claude or Cursor connects to the builder and creates, edits and deploys your apps. Lovable, Replit, Base44, Softr and Webflow ship this. Bolt.new, v0 and Bubble do not.
2. Build-time context. You plug your Notion, Linear or GitHub into the builder's own chat so it reads real context while it works. This is near-universal and it is what most vendors mean when they advertise "MCP".
3. Runtime. The app you shipped becomes an MCP server, so your end users point their own assistants at it. Only Lovable, Base44 and Softr do this.

Only three of the eight builders here do all three, and they are Lovable, Base44 and Softr. That is not the ranking most developers would guess: Bolt.new and v0, the two most developer-facing names in the group, are the two that an agent cannot drive at all.

The practical rule: direction 2 is table stakes and tells you nothing. Direction 1 is what people mean when they say "an agent built my app". Direction 3 is the only one that changes what your customers can do.

One column, three products

Our AI app builder comparison matrix scores MCP and API automation as criterion 6. It is the column that has aged worst, not because the values are wrong but because a single cell cannot hold three independent facts.

Lovable is the one vendor that says this out loud. Its own documentation carries a table headed "How this differs from Lovable's other MCP features", and states plainly that "Lovable uses MCP in three different ways, each connecting a different system for a different purpose". It then separates agent integrations (assistants use your app) from the Lovable MCP server (external tools build in Lovable). Everyone else leaves the reader to work it out.

So we did the work across all eight builders. The result is the grid below, and every cell in it comes from the vendor's own docs, not from a roundup.

Direction 1: can an agent actually build the app for you?

This is the direction people mean when they say an agent built their app, and it is the one that genuinely separates the field.

Lovable exposes itself at https://mcp.lovable.dev. Its docs describe the flow exactly: an agent calls create_project, Lovable builds and returns a preview URL, and you publish with deploy_project. The server "is available on all plans", though on Enterprise workspaces third-party MCP client access is disabled by default and an admin has to turn it on.

Replit runs a server at https://replit-mcp.com/server/mcp over Streamable HTTP with "OAuth using protected-resource discovery", plus native connectors for ChatGPT, Claude and Slack. It documents eight tools, including create_app_from_prompt, update_app_using_prompt, publish_app and get_publish_status. Its docs also tell you not to build your own OAuth server, because Replit publishes the metadata your client needs.

Base44 runs https://app.base44.com/mcp and is the only builder in this group that puts its control plane behind a paywall: "You need a Builder plan or higher to connect an AI assistant through the Base44 MCP server." In exchange you get the deepest toolset of the five, including write_file, edit_file, grep and run_command against the app's sandbox, so the assistant can write the code itself rather than describe a change for the editor to make.

Softr runs https://mcp.softr.io/mcp and is the only one publishing per-client OAuth client IDs for Claude, Cursor, ChatGPT and Mistral, with a personal access token path for anything else. Its docs note the tokens carry an explicit expiry you pick: never, 1 year, 90 days or 30 days.

Webflow ships a remote MCP server over OAuth that creates elements, styles and variables on the canvas and manages collections, assets and custom code. It comes with the sharpest caveat in this entire comparison, and it is in the next section.

Bolt.new, v0 and Bubble ship nothing here. Bolt's only MCP page is about connecting to other servers. v0's docs are explicit that "MCP servers only provide tools for v0 to use during generation"; its programmatic control plane is a plain REST API at api.v0.dev. Bubble's entire manual returns zero matches for MCP; it has a Data API and a Workflow API and that is the whole story.

The Webflow caveat: half of it is not headless

Webflow's MCP server splits into Designer API tools and Data API tools, and only one half runs unattended. From Webflow's own docs: "Designer API calls are executed through a companion app that automatically installs to your authorized sites after OAuth authorization. The companion app must remain open in the Webflow Designer for Designer API tools to function."

The Data API half is fine: "If you're only using the Data API tools, you can use the MCP server to update content even if Webflow isn't open in the browser."

Read that as a straight operational rule. If your plan is a nightly agent that restyles pages, a human's browser has to be sitting open on the Designer for it to work. If your plan is an agent that updates CMS content, it runs headless like anything else. The MCP server is one product with two very different automation stories, and no comparison table we have seen anywhere flags it.

Direction 2: the one everybody has, and the one that proves nothing

Almost every builder lets you connect outside MCP servers into its chat so the model reads your real context while it builds.

Bolt.new offers built-in connectors for Notion, Linear and GitHub plus custom servers, and accepts API key, MCP OAuth, or no authentication for public servers. Lovable calls these chat connectors and adds workspace-level MCP registries so members can browse a directory instead of pasting URLs. Replit ships a curated one-click list. v0 supports bring-your-own servers with four auth modes plus tool calls from Vercel Marketplace integrations. Base44 lets you attach custom MCP servers for the AI chat to use while you build.

This is a real convenience and it is worth having. It is also the weakest possible evidence that a platform is agent-ready, because it describes what the vendor's model can read, not what your agent can do. When a landing page says "MCP support" with no further qualification, this is usually what it means.

Direction 3: your app becomes the MCP server

This is the direction with the most upside and the fewest vendors, and it is the one that changes the product you ship rather than the way you build it.

Lovable generates and hosts an MCP server for a published app, proposing tools based on what the app does. The scoping is clean: "The assistant gets access only to the tools you publish, not to your Lovable project." It is available on all plans for publicly published apps, and sign-in is the default for tool access.

Base44 offers two access modes and the difference matters. Public with no sign-in is read-only, where assistants "can view records from the entities you make available, but cannot change records, message agents, or run custom tools". With OAuth, the assistant "works as the signed-in user" and Base44 "checks those permissions on every request".

Softr has the most interesting mechanism in the whole comparison, because the agent's power is derived from the app's own screens rather than configured separately. A list or table block gives the assistant read access, a form gives create, a profile block gives update, a delete action gives delete. And the boundary is absolute: "If there's no block for a table, the assistant can't touch that table, even if the underlying data source contains it."

That is a genuinely different security posture. On Softr, you cannot accidentally over-expose data to an assistant without first building a screen that exposes it to humans.

The question nobody asks: what does the connection actually reach?

Every comparison stops at yes or no. The more useful question is what a connected assistant can touch on a bad day, and the five control-plane vendors answer it very differently.

Lovable is the outlier and it documents this itself, in a warning block, before the setup steps. "Scope is your full account, not one project. Whatever client you connect can list, read, and edit every project you have access to in Lovable." It goes further on the database tool: "query_database runs SQL with your full database permissions. Read, write, and schema changes." And it reminds you that "Calls run live on your account. Tool calls use real credits and edit real projects."

None of that is a scandal, and Lovable deserves credit for stating it plainly rather than burying it. But it is the single most decision-relevant fact in this entire comparison, and it is invisible in any yes-or-no MCP column, our own included. If you connect a shared or experimental agent to Lovable, it has your whole account and your production schema.

Base44 sits at the other end: the workspace is chosen during OAuth consent and "the connection cannot reach any other workspace", with membership re-checked on every request and token refresh. Softr lets you choose permissions on its own authorization screen and explicitly does not let the tool ask for them in advance: "The AI tool cannot ask for specific permissions in advance. You always choose them yourself on Softr's authorization screen."

Verdict, August 2026

If you want an agent to build and deploy for you: Lovable and Replit are the smoothest, and Replit's OAuth scoping is the more conservative of the two. Base44 is the most capable if you are on a Builder plan, because sandbox file and shell tools let the agent write real code rather than send prompts.

If you are automating an existing site: Webflow, with eyes open. CMS and content automation runs headless. Design automation does not, because the companion app has to stay open in the Designer.

If you want to ship an app your customers drive from their own assistants: Softr, Base44 or Lovable, and only these three. Softr's block-derived permissions are the safest default because the assistant can never exceed what the app's screens already expose.

If MCP is on your requirements list, cross off Bolt.new, v0 and Bubble now. All three are perfectly good builders and two of them have real REST APIs, but no agent is going to drive them over MCP in 2026.

And if you are connecting anything to Lovable, read its scope warning first. Full account access plus SQL with schema rights is a reasonable design, but it is not what most people assume they are agreeing to.

One honest caveat on all of the above: this compares documented capability as of August 2026, verified against each vendor's own docs on August 23, 2026. This category rewrites itself monthly. Two of the cells in our own matrix were already out of date when we ran this comparison, which is exactly why we date every claim.