On this page 6 sections
The Bubble MCP Claude Code setup takes one terminal command, and once connected your agent can edit pages, workflows, data types and option sets in your Bubble app. It’s free while in beta and uses none of your plan’s AI credits. It can’t merge branches or deploy, so a person still reviews and ships every change.
Bubble launched the MCP beta on October 6, 2026, in the same announcement as Agent 2 and AI credits. That timing matters for founders: building with Bubble’s own Agent is now metered, the web-only Starter plan rose to $39 a month on October 20, and for the first time you can bring the coding agent you already pay for. This piece covers what that agent can touch, where it stops, and whether any of it changes the stay-or-leave decision.
What the Bubble MCP can edit
Bubble’s support article on the MCP lists the scope. Through it, an agent can:
- Edit pages, elements, workflows, data types, styles, option sets, API Connector calls and plugins.
- Create, update and delete records, in the development database only. The live database is read-only.
- Run Bubble’s issue checker and query server logs. The October 6 launch post said the beta couldn’t read logs, and the support article now lists a logs tool, so the scope is moving week to week.
- Add sample data and refactor existing structure.
- Create savepoints, restore them, and undo its own steps.
The server acts as your logged-in Bubble account, so the agent can do what you can do and no more. Plan limits still apply: if your plan doesn’t include backend workflows, the agent can’t create them either.
What it can’t do matters more for planning. It can’t merge branches, though you can point it at a specific branch. It can’t deploy. It can’t finish an OAuth2 setup, which stays in the editor, and it can’t install a paid plugin you haven’t bought. Every change sits in development until someone opens the editor and deploys it.
The Bubble MCP Claude Code setup in two commands
The setup is short. In a terminal:
claude mcp add --transport http --scope user bubble https://mcp.bubble.io/mcp
Then open Claude Code, run /mcp, select bubble and authenticate. A browser window opens for the Bubble sign-in. The --scope user flag makes Bubble available in every Claude Code project on your machine. If you’d rather keep it to one repo, the Claude Code MCP docs cover project scope, which writes the server to a .mcp.json file you can commit and share with your team.
For Codex it’s codex mcp add bubble --url https://mcp.bubble.io/mcp followed by codex mcp login bubble. Bubble’s bubble-for-agents repository also lists plugins for ChatGPT, Grok Build and Cursor. One caution from the support article: the server signs clients in with a method some tools don’t support yet, so confirm yours connects before you plan around it.
Two settings worth knowing on the Claude Code side. The docs say Claude Code warns when an MCP tool returns more than 10,000 tokens and caps output at 25,000 by default, which a large Bubble app can hit when the agent reads its structure. MAX_MCP_OUTPUT_TOKENS raises the cap. The docs also tell you to trust a server before connecting it, which is a fair reminder that this one has write access to your app.
Agent 2 or your own agent
Bubble now gives you two ways to build with AI, and they bill differently.
Agent 2 lives in the editor. Bubble says it’s 1.5X better at building what the user asked for and 2X better at producing something that works, and it runs the issue checker in the background while it builds. It draws on your plan’s AI credits: 500 a month on Free, 1,000 on Starter, 2,000 on Growth and 4,000 on Team. Included credits reset on the first of each month and don’t roll over. Extra credits cost $10 per 1,000 on Bubble’s AI pricing page and expire six months after purchase.
The MCP costs Bubble nothing during the beta. The AI work counts against your Claude or ChatGPT plan instead, and Bubble isn’t adding rate limits of its own while the beta runs.
Bubble’s own advice leans toward Agent 2. In Bubble’s September founder AMA recap, Emmanuel recommended that builders keep using the Agent inside the editor, because it has more context on the app than an outside tool. Early builder reports point the other way for larger jobs. One detailed Bubble Forum review from October 12 found the MCP with a strong model could automate “weeks of Bubbling in hours”, while the built-in Agent struggled beyond simple point changes. The same review found the MCP spent its first 20 minutes rendering the site from scratch when there was no repo of notes to read, and couldn’t yet reorder option sets or edit conditions.
Read those as one builder’s beta notes, not a benchmark. The pattern is still useful: the MCP pays off on bulk, repetitive edits by someone who keeps notes for the agent, and Agent 2 is simpler for small, contained changes.
The risks that come with write access
The MCP’s limits double as its safety rails, and a few of them catch people out.
- Editor undo won’t save you. MCP changes don’t go into the editor’s undo history. Restoring a savepoint replaces every change made since, including edits a teammate made by hand in the editor.
- Deleted records stay deleted. Records the agent deletes in development can’t be restored by undoing.
- The issue checker isn’t a full net. A Bubble Forum thread on agent-written expressions from October 16 reports an agent writing an expression that referenced an option set that didn’t exist, and the issue checker didn’t flag it.
- Shared logic breaks quietly. An app where several features share one data type or one backend workflow is where a confident bulk edit does the most damage.
Bubble’s support article suggests trying the MCP on a test app first. We’d go further: work on a branch, take a savepoint before each session, and have someone who knows Bubble read the diff in the editor before anything is deployed.
How to decide what to do next
- Small edits, no developer on hand: use Agent 2 and watch the credit meter. The included allowance covers light use.
- Bulk or repetitive changes, and someone who can review them: use the MCP with Claude Code or Codex on a branch. It’s free during the beta, and the editor work stays reviewable.
- Paying for Agent top-ups every month: compare that line with what your existing model subscription already covers. The MCP is free while the beta lasts.
- Workload bills, integrations or page speed are the problem: an AI agent doesn’t change those. Our guide to when to rebuild your Bubble app covers the four thresholds that decide it.
- No one on the team to review agent changes: hire Bubble developers who can run the MCP on a branch and check each change before it ships.
- Already planning to leave: the MCP can read your app’s structure for the inventory, but the data still moves the long way. See our walkthrough on migrating a Bubble app to Supabase.
The Starter price change sits under all of this. A founder on web-only Starter now pays $39 a month, or $32 billed annually, plus any Agent credits past the included 1,000. That’s still cheap next to a rebuild, which is why most Bubble apps should stay put until a threshold forces the question.
How Glibzter shipped on Bubble in 6 weeks
We built Glibzter’s coaching platform on Bubble in 6 weeks: scored assessments, a video and audio course player, calendar-based 1:1 coach booking and Stripe checkout for both courses and consulting. Two product surfaces run on one shared auth, wallet and content backbone in the same Bubble app.
On an app built that way, the shared backbone is what an agent with write access should treat with care. A change to the wallet or the assessment scoring touches both products at once. Here, the time an AI agent saves on page and workflow edits is useful, and the review step before deploy is the part you don’t skip. When a Bubble app reaches the point where those shared pieces need to live in code, our Bubble to code migration moves them one at a time, and our Code Rescue engagement cleans up apps an agent or a rushed build has left tangled.
Frequently asked questions
- Is the Bubble MCP free to use?
- Yes, while it's in beta. Bubble says the MCP is free during the beta, works on every plan and doesn't draw on your AI credits. The AI work is billed by your model provider instead, so a long Claude Code session counts against your Claude plan's usage limits. Bubble isn't adding its own rate limits during the beta.
- Can Claude Code deploy my Bubble app through the MCP?
- No. The Bubble MCP can't deploy, merge branches, finish OAuth2 setup or install paid plugins you haven't bought. It edits the development version and can only read the live database. You review the changes in the Bubble editor and deploy them yourself, which keeps a human between the agent and your users.
- Should I use Bubble Agent 2 or my own agent through the MCP?
- Use Agent 2 for small, scoped edits inside the editor, and the MCP for bulk or repetitive work you'd otherwise click through by hand. Agent 2 draws on your plan's AI credits, from 500 a month on Free to 4,000 on Team. The MCP draws on your own model subscription and needs someone who can review what it changed.
- Can I undo changes the Bubble MCP made?
- Partly. MCP changes don't go into the editor's undo history, so the editor's undo button won't reverse them. You can create and restore savepoints through the MCP, but restoring one wipes every change made since, including your own edits in the editor. Records the agent deletes can't be brought back by undoing, so test on a copy first.
- Does the Bubble MCP work with Codex and Cursor?
- Codex is supported with two terminal commands, and Bubble's own repository lists plugins for ChatGPT, Codex, Grok Build and Cursor. Bubble's support article warns that some MCP clients don't support the sign-in method its server uses yet, so check that your tool connects before you plan a project around it.
References
- Bubble - Introducing Agent 2 and the Bubble MCP
- Bubble support - Bubble MCP
- Bubble - Bubble MCP overview and FAQ
- Bubble support - AI credits
- Bubble - AI credit pricing
- Bubble - Bubble for Agents repository
- Claude Code docs - Connect Claude Code to tools via MCP
- Bubble - September demos and AMA recap
- Bubble Forum - The good, the bad, and the amazing (MCP/Agent feedback)
- Bubble Forum - Illegal expressions that AI agents are managing to write in Bubble