Built-in MCP server

Your customers' AI can do the busywork. A person still approves it.

Every ChaiBuilder deployment includes a built-in MCP server at /api/mcp. Your onboarding team and your customers can connect Claude, Cursor, VS Code, or any MCP client to find, edit, translate, and audit pages using their exact user permissions. Every change saves as a draft for human review, with no third-party cloud in between.

Book a demo Read the MCP setup guide

Ask for the change. Review the draft. Publish.

Real tasks your onboarding team and customers can hand to an AI agent on a ChaiBuilder site.

“Build a seasonal catering page from this brief using our custom menu blocks.”

The agent creates the page, populates your registered custom React blocks, and returns a draft preview link for review.

“Find every page on this site with a missing meta description and draft the fixes.”

The agent lists the site's pages, audits their SEO fields, and saves the missing metadata as drafts.

“Translate the Services and About pages into Spanish and German.”

For any language enabled on the site, the agent reads the translatable blocks and writes the localized drafts ready for a human eye.

“Pull this week's featured listings using our custom inventory tool and update the homepage hero.”

Using a custom MCP tool registered by your developers, the agent reads live records from your platform database and updates the page draft.

“We moved /services/dental-implants to /implants. Set up the redirect.”

The agent creates the URL redirect and confirms it is active immediately.

“What changed on the Pricing page since Tuesday? Restore the previous version.”

The agent compares page revisions, summarizes the diff, and restores the selected version.

Built so an agent can never go further than the user who connected it

An agent is not a superuser. It runs through the exact same permission checks, draft workflow, and revision history as the visual builder.

Scoped to the user's exact permissions

Each user connects with their own personal API key. Every MCP tool requires the same RBAC permission as the matching action in the visual builder, and tools the user is not allowed to run are hidden from the agent.

Dedicated permission to connect (mcp:use)

Connecting an AI agent is gated by its own permission key. You can let a customer edit visually in the builder without allowing them to connect an external agent.

Drafts first, human approval to publish

Content, block, SEO, and translation edits always land in the page's draft state with a preview link, and the agent is instructed to ask before publishing.

Real-time collision lock with human editors

While an agent edits a page, the visual builder marks that page "AI is editing" and opens it read-only. Conversely, if a person already has a page open in the builder, the agent's write is refused so no one ever loses work.

Opt-in Design System for Agents

Enable designSystem on the MCP plugin to check every block an agent writes against your site's typography, color, and layout rules before it lands.

Traceable revisions and onToolCall audit hook

Every agent edit is tagged in page revision history for one-click rollback, and your app can log every tool call (user, page, tool, duration, and outcome with secrets redacted) via the onToolCall hook.

Expose your platform's own data and actions to AI agents

The 33 built-in MCP tools handle pages, blocks, SEO, translations, redirects, and revisions. Your engineers can register custom MCP tools right alongside your custom React blocks.

Read the custom MCP tools guide

Query your platform database

Let an agent look up a customer's live inventory, property listings, staff bios, or appointment availability while drafting a page.

Wrap your existing server actions

Expose validated backend actions you already wrote for the visual builder so agents can invoke them under the same permission checks.

Encode industry-specific playbooks

Provide custom agent instructions and design rules tailored to your vertical so every page an agent drafts follows your platform's best practices.

Your agent talks directly to your deployment. Not to our cloud.

A hosted builder's MCP server

Your customer's agent connects to a third-party vendor's cloud, under the vendor's rate limits and terms, to edit content stored in the vendor's database.

An AI code generator

An agent writes raw source code that still has to be reviewed, compiled, built, and deployed through CI/CD before a single text change goes live.

ChaiBuilder's self-hosted MCP server

The MCP endpoint (/api/mcp) runs inside your own Next.js deployment and writes structured JSON drafts directly into your own PostgreSQL or SQLite database. When approved, publishing is live instantly with no build or deploy.

Learn more about data ownership

Any MCP client. Zero AI credit fees from us.

Supported clients

Claude Desktop

Claude Code

Cursor

VS Code (Copilot agent mode)

Any client that supports Streamable HTTP

Zero token markup

Because the model runs inside the user's own AI client and talks directly to your self-hosted /api/mcp endpoint, ChaiBuilder charges no AI credits and no per-agent fees. The MCP server is included in every Pro license.

Opt-in by design. Enable once, connect in a minute.

01

Your developer enables the /api/mcp route

Because MCP lets external agents write drafts to your sites, it is opt-in: register the MCP plugin, enable user API keys, and add the route handler once in your Next.js app.

02

Authorized users generate a personal API key

Any user granted the mcp:use permission creates an API key in their profile. That key carries their exact identity and site permissions.

03

Connect Claude, Cursor, or VS Code

Paste the site's MCP URL and personal API key into the AI client and start drafting pages.

Read the step-by-step setup docs

Questions about the MCP server

Is the MCP server included in the Pro license?

Yes. Every Pro license includes the built-in MCP server, all 33 built-in tools, custom MCP tool registration, and the opt-in Design System for Agents at no extra cost.

Can an AI agent publish changes to a live customer site without approval?

Edits to page content, SEO, and translations always save as a draft first, and the agent is instructed to return a preview link and ask for human confirmation before publishing. If you want to prohibit agents from publishing altogether, simply withhold the publish permission from that user role or omit the publish tool in your config.

Is the MCP server enabled automatically on existing projects?

It is enabled in new starter deployments, and on existing projects a developer switches it on once in code so you stay in full control of when agent access is opened.

Where does page data go when an agent reads or edits a page?

Directly between your self-hosted /api/mcp endpoint and the AI client the user connected (such as Claude or Cursor). ChaiBuilder never sits in the middle or sees your traffic.

Does the MCP server run on serverless hosts like Vercel or Cloudflare?

Yes. The MCP endpoint is stateless and runs cleanly on serverless Next.js hosts as well as standard Node or Docker servers.

See AI agents and human editors working on the same customer site.

Book a 30-minute demo to see the built-in MCP server, custom data tools, and draft-approval workflow in action.

Book a demo Compare with the in-builder AI chat assistant