Bubble MCP and Agent 2: What Bubble's New AI Tools Do and How to Connect Them
By Ethan Halfhide · Updated October 12, 2026
Bubble launched Agent 2 and a beta MCP server on October 6, 2026. Here is what each one does, how to connect your own AI tool, and what it costs.
Bubble Agent 2 and MCPOn October 6, 2026, Bubble shipped two AI releases at once. Agent 2 is a faster, more accurate version of the Bubble AI Agent that lives inside the editor. Bubble MCP is a new server, in beta, that lets you build and edit Bubble apps from your own AI tool, such as Claude, Claude Code, ChatGPT or Codex. The same announcement introduced AI credits for the Agent and a higher price for the web-only Starter plan.
This guide explains what each release does, how they differ, how to connect your own AI tool to Bubble, and what it costs. Everything here comes from Bubble's announcement and its official manual, read on October 11, 2026. We have not run either tool, so there are no test results on this page.
Quick facts (as of October 11, 2026)
| Agent 2 | Bubble MCP | |
|---|---|---|
| What it is | Bubble's own AI agent, inside the editor | A server that connects your own AI tool to Bubble |
| Status | Out of beta, available to all users on all apps | Beta |
| Where you work | The Bubble editor | Claude, Claude Code, ChatGPT, Codex, or VS Code with GitHub Copilot |
| Which AI model | Bubble's choice | The model you pick in your own tool |
| Cost | Uses Bubble AI credits | Does not use Bubble AI credits. Usage counts toward your AI provider plan |
| Live data | No live database access. Works in the development database only, and reads records only for owners and admins, for data types enabled in settings | No live database access. Reads and writes the development database only |
| Deploys your app | Not stated by Bubble | No, you deploy |
What is Agent 2?
Agent 2 is a major upgrade to the Bubble AI Agent, the assistant pinned to the right side of the Bubble editor. It can generate a new app from a description, then keep editing it: adding pages, wiring workflows, structuring the database, connecting external APIs and changing styles. It works on both web and native mobile apps.
Bubble says Agent 2 is 1.5 times better at building exactly what the user asked for and 2 times better at producing a fully functional result. Part of the reason, according to Bubble, is that it runs the editor's issue checker in the background while it builds and learns from what it finds. Bubble also says it handles vague prompts better and can manage larger edits, such as adding a full feature to an established app. These are Bubble's own figures. Bubble did not publish the test method behind them, and we have not tested them.
The workflow inside the editor changed too:
- Preview is now its own tab in the editor. When an app finishes generating, you land on the preview with the Agent pinned beside it.
- Edit while you preview. You can ask for changes from the preview tab and watch the app update.
- Point at what you mean. Click a button, a workflow step or a data field to focus the Agent on that exact piece.
- Run as a user. A dropdown in the preview tab lets you view the app as a specific user.
Bubble notes that Safari support for the new preview is still in progress.
The Agent follows your own permissions. If you have editor access, it can change the app on your behalf. If you have view-only access, it can inspect and explain the app but will not edit it. Bubble's manual says the Agent works only with your development database, never the live one, and reading records there is limited further, to owners and admins and only for the data types enabled in Agent settings.
What is Bubble MCP?
MCP stands for Model Context Protocol, an open standard for connecting AI tools to other software. An MCP server is the connection point a service offers. Bubble's server lives at https://mcp.bubble.io/mcp. Once your AI tool is connected, you can ask it to create a new Bubble app or change an existing one, and it builds directly in your Bubble account.
The key point is that the result is still a normal Bubble app. Your AI tool builds with Bubble's own elements, workflows and data types, not generated code, so you can open the app in the visual editor and change anything by hand afterward.
Bubble positions the MCP for experienced builders who want to edit and grow apps they already have from their own agent setup. Its announcement recommends trying Agent 2 first.
What the MCP server can do
Bubble's manual lists the server's tools in 13 groups:
| Group | Examples from Bubble's tool list |
|---|---|
| Apps and account | list_apps, create_app, whoami |
| Exploring an app | get_app_outline, get_pages, get_app_plan_context |
| Pages and elements | create_page, create_element, set_element_properties |
| Styles and variables | create_style, create_color_token, create_font_token |
| Workflows | create_event, create_action, reorder_actions |
| Expressions | validate_expression, create_global_expression |
| Data types and privacy | create_data_type, create_field, create_privacy_rule |
| Records and files | search_things, bulk_create_things, upload_sample_data |
| Option sets | create_option_set, create_option |
| API Connector | create_api_collection, create_api_call |
| Plugins | search_plugins, install_plugin |
| Testing, preview and issues | get_preview_link, check_app_issues, create_test_cases |
| Changes, undo and savepoints | apply_changes, undo, create_savepoint |
In practice, Bubble's manual and announcement describe uses like these:
- Create a new app from a description.
- Add pages, elements, workflows and data types to an existing app.
- Find out why something is broken and fix it.
- Refactor, for example by combining workflows that do the same thing.
- Fill the development database with sample records.
- Set up backend workflows and API calls.
- Pull context from other MCP servers you have connected, such as a Linear or Jira ticket, a Figma mockup or a spec in your documents, and build from it.
What it cannot do yet
- Deploy your app. Changes only reach your live app when you deploy them yourself in the editor, which gives you a chance to review them first.
- Merge branches. You can tell the agent which branch to work on.
- Read logs.
- Set up OAuth2. That stays in the editor.
- Install paid plugins you have not bought yet.
Permissions and data access
The MCP server acts as your signed-in Bubble user, so your AI tool can only do what you are allowed to do. Collaborators who connect their own tools work within their own permissions. Your app's plan still applies too: if the plan does not include backend workflows, the AI tool cannot create them.
Data access is limited to the development database:
| Database | Read | Write |
|---|---|---|
| Development | Yes | Yes |
| Live | No | No |
Record tools also only work for the data types an admin has granted the AI tool access to.
Undo works differently
This matters before you let an agent loose on a real app. Changes made through the MCP server do not appear in the Bubble editor's undo history, so pressing undo in the editor will not reverse them. Instead you can ask the AI tool to undo, including everything from one conversation, or ask it to create a savepoint before a big change and restore it later. Restoring a savepoint replaces every change made since, including edits you made by hand, and how far back you can go depends on your plan. Records the AI tool deletes cannot be restored by undoing. Bubble suggests trying the server on a test app first.
How to connect your AI tool to Bubble MCP
Every client follows the same pattern: add the server, then sign in to your Bubble account in the browser and approve the permissions. Bubble's server uses CIMD (Client ID Metadata Documents) for sign-in. Bubble says Cursor and Composio do not support CIMD yet, so they cannot connect for now.
Claude: go to Customize, then Connectors. Click +, search for Bubble MCP, click Connect and sign in to Bubble. In a chat, click +, select Connectors and turn Bubble MCP on.
Claude Code: run this in your terminal, then open Claude Code, type /mcp, select bubble and authenticate in the browser.
claude mcp add --transport http --scope user bubble https://mcp.bubble.io/mcp
The --scope user part makes Bubble available in all your projects. To limit it to one project, add a bubble entry with "type": "http" and the server URL to .mcp.json in the project root instead. If you use Claude Code, Bubble's manual walks through this setup step by step.
ChatGPT: on a personal account, turn on Developer mode under Settings, then Security and login. On Enterprise or Edu workspaces, an admin must enable Developer mode access. Then go to Plugins, click +, choose Create custom MCP server and Create MCP app, name it Bubble, paste the server URL, keep authentication set to OAuth, confirm and click Create.
Codex:
codex mcp add bubble --url https://mcp.bubble.io/mcp
codex mcp login bubble
VS Code with GitHub Copilot: add this to .vscode/mcp.json, click Start above the entry and sign in.
{
"servers": {
"bubble": {
"type": "http",
"url": "https://mcp.bubble.io/mcp"
}
}
}
Once you are connected, describe what you want in plain language. If you have more than one app, name the app in your prompt so the agent works on the right one. Bubble's manual gives examples like "Create a new Bubble app for a book club, where members can suggest books and vote on what to read next" and "Create a savepoint before you start, then rebuild the checkout page."
Pricing: AI credits and the Starter change
Agent 2 uses AI credits. Since October 6, 2026, building and editing with the Bubble AI Agent uses AI credits, and so does generating images. Generating a new app, including its planning session, does not use credits, and neither does editing in the visual editor or building through the MCP. Bubble's manual also lists generating test data as a credit use.
Every plan includes a monthly allowance. Bubble's manual lists these amounts as of October 11, 2026, the same for web, mobile and combined plans and for monthly or annual billing:
| Plan | Included AI credits per month |
|---|---|
| Free | 500 |
| Starter | 1,000 |
| Growth | 2,000 |
| Team | 4,000 |
| Agency | 1,000 |
| Enterprise and Dedicated | Custom |
A few details worth knowing:
- Credits belong to the app, not your account. Each app has its own balance, and every collaborator on that app draws from it. You cannot pool credits across apps.
- Included credits reset on the first of each month and do not roll over. Credits you buy last six months.
- There is no fixed cost per message. Bubble calculates credits after a request runs, and big changes that touch many elements use more, even from a short prompt.
- Running out does not charge you. The Agent stops, keeps the work it finished and summarizes what is left. You can keep building in the visual editor, buy more credits on a paid plan, upgrade, or wait for the monthly reset. Free plan apps need to upgrade to Starter to buy more.
- Launch bonus. Apps that existed on October 6, 2026 get double their plan's credits through October 31, 2026. Apps created after that date, new accounts, trials and apps that change plans in October do not get the bonus.
The MCP does not use Bubble AI credits. The AI work happens in your own tool, so it counts toward your plan with that provider, such as your Claude plan, and you may hit that provider's rate limits.
Web-only Starter price change. Starting October 20, 2026, the web-only Starter plan rises to $39 a month on monthly billing or $32 a month billed annually. Current Starter Web and trial users move to the new price at their next renewal. Bubble says you can lock in the current $29 a month annual rate for a year by buying the annual plan before October 20. Check Bubble's pricing page for current prices, since Bubble says that page is the authoritative source.
Agent 2 or Bubble MCP: which should you use?
Start with Agent 2 if you are new to Bubble, want to build inside the editor, or want the app generated and edited without any setup. It is the option Bubble recommends trying first.
Use the Bubble MCP if you already work in Claude Code, Codex, ChatGPT or Claude every day, you want to choose the model, or you want your agent to pull from other tools like your ticket tracker, design files or documents while it builds. It also suits builders who would rather spend their AI provider's usage than Bubble AI credits.
Wait or plan around it if you rely on Cursor or Composio, which cannot connect yet, or you need the agent to deploy, merge branches or read logs. You will still do those in the editor.
You can use both. Bubble's manual describes using the Agent while you work in the editor and the MCP when you want to work from another agent or try a different model.
How this changes Bubble's place among AI app builders
Until this release, building with AI on Bubble meant using its own in-editor Agent. In its October update, Bubble describes the MCP as "opening Bubble up to every agent." It brings Bubble to the coding agents many builders already use, while keeping the thing that sets Bubble apart: the result is a visual app you can inspect and edit without reading code. That is a different trade-off from AI builders that generate code you then own and maintain, such as Lovable. For a side-by-side look at the wider field, see our guide to the best AI app builders, or read our Bubble listing.
FAQ
Is Bubble MCP free? Bubble does not charge AI credits for the MCP, and Bubble's forum announcement says there are no request limits, including on the Free plan. The AI work counts toward your plan with your AI provider, so the real cost depends on that provider.
Can Bubble MCP change my live app or live data? No. It has no read or write access to the live database, and it cannot deploy. Changes stay in development until you deploy them yourself.
Does Bubble MCP work with Cursor? Not yet. Bubble's server uses CIMD sign-in, which Bubble says Cursor and Composio do not support yet. Bubble says it is working on support.
Do I need to code to use it? No. You describe what you want in plain language and the result is a visual Bubble app. Connecting Claude Code or Codex means pasting a command or two into a terminal, which Bubble's manual walks through.
Is Agent 2 available on existing apps and mobile apps? Yes. Bubble says Agent 2 is available to all users on all apps, including existing apps, for both web and native mobile.
What happens when I run out of AI credits? The Agent stops and keeps the work it finished, and no automatic charge happens. You can keep editing in the visual editor, buy more credits on a paid plan, upgrade, or wait for the reset on the first of the month.
Sources
- Introducing Agent 2 and the Bubble MCP, Bubble blog, October 6, 2026
- Bubble MCP, Bubble manual
- AI credits, Bubble manual
- Bubble AI Agent, Bubble manual
- Monthly Community Update, October 2026, Bubble forum
- Bubble pricing
About the author
Ethan Halfhide
Ethan Halfhide is an AI technologist and founder who has helped more than a thousand people build websites and products with AI tools. He compares AI website builders on pricing, features and what each one actually produces from a prompt.