Anticode as a ChatGPT website builder connector: research, build, publish
ChatGPT can reach Anticode through its remote MCP server at https://anticode.dev/mcp. Once connected with OAuth, it can look through your projects with the read-only search and fetch pair that deep research uses, and, if you granted write access, create projects, run quoted builds, request changes and publish. What you can add depends on your ChatGPT plan and workspace settings.
Adding the server in ChatGPT
OpenAI has moved this setting more than once, so follow whichever screen your account shows. Anticode's own setup page at anticode.dev/mcp currently describes the developer-mode path: in ChatGPT's settings, open Apps & Connectors, then Advanced settings, turn on Developer mode, then choose Create, name it Anticode, paste the server URL and pick OAuth. OpenAI's current developer guide to connecting an MCP server to ChatGPT currently describes the same job from the plugins page:
- Choose the plus button, then Add custom MCP server.
- Give it a name and a description, and under Connection enter the public endpoint
https://anticode.dev/mcp, including the/mcppath. - Configure authentication (OAuth), read the risk warning and confirm.
- Review the tools ChatGPT discovered, then finish.
- Approve the connection on the Anticode page that opens. You can narrow it to read-only there.
OpenAI notes that account and workspace policies govern who can add custom servers. In a business workspace an admin may need to allow it first, and plans can differ in what custom connectors are allowed to do. Check your plan's current rules in OpenAI's help centre rather than relying on a summary that may have aged.
Search and fetch: what deep research sees
ChatGPT's research features look for two read-only tools with fixed shapes, documented in OpenAI's MCP guide: search takes a query and returns results with an id, title and URL; fetch takes an id and returns the full text. Anticode implements both over your own projects only:
- search
- Matches words against each project's name, original request, page copy and offerings, best matches first, up to 20. An empty query lists your most recent projects. Each result's URL is the live site if it's published, otherwise your dashboard.
- fetch
- Takes an id such as
project:3f2a…and returns one readable document: the request, kind, design system, theme, sections, headline, offerings with prices, opening hours and place, and where it's published. Metadata adds the kind, design, version, last update and whether it's live.
That makes questions like "which of my sites still says we close at 5 on Saturdays?" answerable across every project, with citations that link to the live pages. Neither tool changes anything, and neither returns form submissions or analytics.
Which tools need write access
| Read-only connection | Also needs read and write |
|---|---|
search, fetch, list_projects, get_project, get_project_file | create_project, update_project_file |
get_build_quote (without a usable quote id) | start_build, request_change |
preview_project, list_sites | publish_project, unpublish_site |
get_analytics, list_form_submissions |
If ChatGPT calls a write tool on a read-only OAuth connection, Anticode answers with an insufficient_scope error, the standard signal a client can use to ask you for more access. Even with write access, a build only runs with a quote id issued in the last 15 minutes for that project and purpose, used once, and the tool tells ChatGPT to show you the cost first.
How the conversation goes
A typical session: you describe a site; ChatGPT creates the project (free and local), asks for a quote, tells you the cost against your daily and weekly allowance, and waits. You agree; it runs the build, then shows the page. Hosts that support MCP Apps can display it inside the chat; elsewhere, ChatGPT reads a summary. You ask for changes, it applies the free ones first, and publishing renders the page with Anticode's design engine and passes it through the security gate.
Limits
- No ChatGPT-specific widgets. Anticode's in-chat preview uses the MCP Apps standard, not OpenAI's own output template format.
- No uploads or deletions of projects, files or submissions through any assistant.
- Search is word matching over your projects, not semantic search across the web.
- Form submissions are untrusted. They're labelled that way, because visitors wrote them.
- Plan rules are OpenAI's. Whether you can connect, and with which tools, is decided by your ChatGPT plan and workspace.
Troubleshooting
- No option to add a custom server
- Your plan or workspace doesn't allow it, or an admin hasn't enabled it.
- Write tools are missing or refused
- The connection is read-only, on Anticode's side or ChatGPT's. Reconnect with read and write where your plan allows.
- Search finds nothing
- Try words from the project's name or headline, or an empty query to list recent projects.
Every tool is described on the MCP server page. If you use both assistants, managing a website from ChatGPT or Claude covers the routine, and what an MCP server is explains the protocol.
Questions people ask
Can ChatGPT read my visitors' form submissions?
Only through list_form_submissions, which needs you to have connected the account, and the results are labelled as untrusted content.
Does deep research see other people's projects?
No. Search and fetch run as your account and only ever return your own projects.
How do I disconnect ChatGPT?
On anticode.dev/mcp, choose Disconnect next to the app. Access ends immediately.