Build and publish from your AI assistant with Anticode's MCP server
Anticode runs a remote MCP server at https://anticode.dev/mcp. Connect it to Claude, ChatGPT, Cursor or another client that supports OAuth remote servers, and the assistant can list and create your projects, quote and run builds, change them, edit the project file, publish and read your analytics and form submissions. Every action goes through the same checks, limits and cost quote as the web app.
What an MCP website builder adds
The Model Context Protocol lets an AI assistant call tools that another service provides. With an MCP website builder connected, you can stay in the chat you already use and say "make a booking page for my studio, show me the cost first" or "what came through the contact form this week?" The assistant does the work through Anticode's tools, and the site that comes out is the same one you'd get from the builder: one of the seven design systems, the Critic's design checks and Sentinel's security gate before it goes public.
Connecting a client
Every client uses the same server address. Clients with OAuth support sign you in through the browser and show Anticode's consent screen, where you can see which app is asking and narrow its access from read-and-write to read-only.
- Claude (claude.ai and Claude Desktop)
- Add a custom connector named Anticode with the server URL, then approve it on Anticode.
- Claude Code
claude mcp add --transport http anticode https://anticode.dev/mcp, then run/mcpinside Claude Code to sign in.- Cursor
- Add the server URL under
mcpServersin Cursor's MCP settings file. - ChatGPT
- Create a connector with the server URL and OAuth authentication from ChatGPT's developer mode settings.
- Other clients
- Any client that supports OAuth remote MCP servers can connect. Clients without OAuth can send a personal token made on the /mcp page instead.
Personal tokens can be read-only or read-and-write and can be revoked on the /mcp page, where you can also disconnect any app you approved. The guides for Claude, ChatGPT and Cursor go through each client in more detail.
The tools an assistant gets
| Tool | What it does | Access |
|---|---|---|
| list_projects, get_project | Find a project and read its overview: request, design direction, sections and copy | Read |
| create_project | Start a project from a one-sentence description, with no model call | Write |
| get_build_quote | What a build or change will cost against the daily and weekly allowance, and whether it can run | Read |
| start_build, request_change | Run a model build, or apply a change described in words | Write |
| get_project_file, update_project_file | Read and replace the project's anticode.json for exact wording, order and theme | Read / Write |
| publish_project, unpublish_site, list_sites | Put a site live behind the security gate, take it down, list what's live | Write / Read |
| get_analytics, list_form_submissions | Page views by day and source; visitor submissions | Read |
| preview_project, search, fetch | Show the page in the chat; search and fetch for ChatGPT | Read |
Nothing in the list uploads files or deletes projects or submissions; those stay in the web app.
The assistant has to show you the cost
A model build or a model-written change only runs with a quote issued for that account, project and purpose in the last fifteen minutes. The quote tool returns a plain cost summary, and its instructions tell the assistant to show you that summary and wait for your go-ahead before passing it on. A quote can be used once. If the build fails, it costs nothing and the quote is handed back. Many changes don't need the model at all: request_change first tries a free local change, which understands direction, theme, action colour, headline scale, adding or removing sections and renaming.
Safety and what stays the same
- Same code path. Each tool call runs the web app's own handlers, so ownership checks, allowances, rate limits and the publish gate are identical to the builder's.
- Tokens, not cookies. The server only accepts bearer tokens, so another website can't act through your signed-in browser.
- Short-lived access. OAuth access tokens last an hour and refresh tokens rotate on each use; disconnecting on /mcp ends the connection at once.
- Submissions are marked untrusted. Form submissions come back flagged as visitor content, so the assistant treats them as data rather than instructions.
- Edits show up in History. Saves made through MCP are labelled "AI assistant (MCP)" in version history, and an open builder redraws when they land.
Limits
- Progress is per step. Clients that ask for progress see steps such as planning and saving, not each word as the model writes it.
- In-chat preview depends on the client. The preview uses the MCP Apps standard; hosts that support it, including Claude, Claude Desktop and VS Code, show it. Others get a text summary.
- CMS sections when publishing. Sections filled from CMS collections aren't rendered when publishing through MCP; the tool says so when a project has them. Publish those from the builder.
- Client plans. Some assistants only allow custom connectors on their own paid plans or in a developer mode. Check your client's rules.
Questions people ask
Can the assistant spend my allowance without asking?
A model build needs a quote the assistant fetched first, and the tool tells it to show you the cost and get your agreement. A read-only connection can't start builds at all.
Does my assistant see my password?
No. OAuth sign-in happens on Anticode's own page in your browser, and the assistant only receives a token you can revoke.
Can I use the MCP server and the builder at the same time?
Yes. Both work on the same project. If someone saves in between, a stale edit to the project file is refused with a version conflict so nothing is overwritten.
Is this the same as the CLI?
No. The CLI works on a project from your terminal; the MCP server is for AI assistants. Both publish through the server's security gate.