Drivethrough MCP Server for ChatGPT
By Made by Tim LLC
Overview
Describe the website you want, and it's live in seconds. No code, no settings, no website builder to learn - you say what you want, ChatGPT builds it, and Drivethrough puts it online. Every site starts private, visible only to you. Share it like a doc link - send it to exactly the people you choose, no accounts needed - or make it public for the world when you're ready. That makes Drivethrough good for two kinds of sites: the little ones that were never for the whole internet (trip plans, the family recipe book, a tool that splits the bills), and the public ones you've been meaning to make (the plumbing business, the band, the bake sale). Changing a site is just asking again. Every change is saved, so you can roll back to any earlier version instantly. Forms and RSVPs land in your email. Public sites can use your own web address, and it's free to start.
How to connect
- ChatGPT: Open the Drivethrough listing in the ChatGPT app directory and choose Connect. ChatGPT listing Drivethrough exposes 21 tools here.
Tools
create_db
ChatGPTcreate_integration_setup_link
ChatGPTcreate_upload
ChatGPTsite plus the returned session — it lists what they uploaded ({upload_id, filename, size_bytes}); ship each with deploy_site files: [{path, upload_id}] (mode "patch" adds them to a live site). Never base64-encode a file instead — but if the bytes are already in YOUR runtime (e.g. a chat attachment mounted in your sandbox) AND you can make HTTP requests, you may PUT them straight to the returned put_url (one file per request, its filename as the last path segment) without involving the user. Choosing a route for a binary asset: already on the public web → deploy_site {path, source_url}; this server runs on the user's machine → deploy the directory; file only in the chat or on the user's machine → this tool. Tiny pasted assets (≤1 MB, e.g. a favicon) may still go inline as base64 content. Upload ids are single-use: the deploy that references one consumes it (the bundle keeps the bytes; the staged copy is removed and its quota refunded) — to rename or duplicate the file later, use deploy_site {path, from_path}. An upload no deploy references within 24 hours is deleted, so deploy soon after the user confirms. Pass size_bytes when the user told you a file's size — quota errors then surface here instead of failing their upload.delete_site
ChatGPTdeploy_site
ChatGPTfiles with generated content — the platform assembles the bundle server-side, so no local filesystem or build toolchain is needed. Content you generated counts even once it is sitting in a file: if you wrote the page to your own sandbox, read that text back and send it as files[].content. Images and other binary assets — pick the route by where the bytes live: already at a public URL → {path, source_url} (fetched server-side); a local file — the user's, or an attachment already in your runtime — → create_upload first, then {path, upload_id}; already deployed on this site → {path, from_path} copies it server-side (with delete_paths, a move). Never base64-encode a file, by hand or in code — inline base64 is capped at 1 MB and is only for tiny pasted assets; a file shipped via upload_id is referenced, never additionally inlined. To rework an already-deployed site, read it back with get_site first, then deploy with mode "patch" — only the files that change. For small text changes to existing files, use edit_site instead: it sends find/replace pairs, never whole files. Every deploy's version number is a restore point: prior versions stay listed (get_site history), diffable, viewable (preview_site), and returnable (revert_site). A successful deploy may still return warnings — things that will bite at runtime but only for visitors, never for you: a form type anonymous visitors can write with no page loading the Turnstile widget, or a page that fetches a data type its visitors have no permission to read. Resolve them before telling the user it works.edit_site
ChatGPTfind must match the deployed text exactly (whitespace included) and exactly once — add surrounding context to pin down one occurrence, or set replace_all. All-or-nothing: if any edit fails to match, nothing deploys and the error says how to recover, so you can try an edit without reading the file first. Edits are never lossy: the pre-edit version stays restorable via revert_site, so to undo an edit, revert — don't hand-invert it. After visual changes, preview_site shows what actually rendered.get_integrations
ChatGPTget_site
ChatGPTfind locates text without returning file contents, manifest_only lists what exists, paths fetches only named files — a bare call inlines every text file, so reserve it for genuinely reading a whole small site. Versions: every deploy is a numbered version the platform keeps. history: true lists them with what each one changed (the map for revert_site), compare: {from, to} returns line diffs between two versions ("what did v9→v10 change?"), and version pins any read mode to a prior version instead of the live one. Modes are exclusive per call — compare wins over history, history over find/manifest_only. Entries come back deploy_site-compatible ({path, content, encoding}); binary files are listed but only inlined when named in paths. For small text changes you may not need this at all — edit_site applies find/replace to the deployed files directly.list_sites
ChatGPTmanage_domain
ChatGPTpreview_action
ChatGPTpreview_schedule
ChatGPTpreview_site
ChatGPTversion before reverting to it. Works on private sites; nothing becomes public. Takes a few seconds — look once after a coherent batch of changes, not after every edit. Animations: the capture happens at page load; pass wait_ms to let one settle or reach mid-cycle.preview_webhook
ChatGPTquery_db
ChatGPTseq is the row number and its permanent identity (never reused, even after a delete), json is the document. Newest-first is the default and always free. Filtering, sorting and search need a matching entry in the site's "query" block in drivethrough.json — an undeclared query is refused with the exact line to add. As the owner you get a bounded 5,000-row scan instead of a refusal, marked "partial": true, so you can look at data you have not modelled yet; a visitor page gets the refusal.restore_site
ChatGPTsitename to restore under a new name. (Undoing changes to a live site is different — that's revert_site.)revert_site
ChatGPTset_privacy
ChatGPTshare_site
ChatGPTwrite_db
ChatGPTseq, the row number query_db returns. To change a few fields of a row, update with merge — only the fields that change, no fetch needed. Reserve json (wholesale replacement) for rewriting the whole document. To load a catalogue, use rows for a handful and rows_url for thousands — never loop insert, and never try to emit 100,000 rows yourself. Writes are validated against the types declared in the site's drivethrough.json (unknown types and mismatched fields are rejected, and a bulk import is checked in full BEFORE the first row is written, so it is all-or-nothing); pass validate: false only for a deliberate migration. Deleted and overwritten rows are gone for good — there is no version history for data, so read a row (query_db) before destructive rewrites you might need to undo. Deleting a row also deletes its children. Every write counts against the account's monthly write allowance, and an import larger than what remains is refused before anything is written.Capabilities
Example Prompts
Click any prompt to copy it.
Links
Similar Apps
Listed on MCP App Store
Add this badge to your site or README. It links to this page.
HTML
<a href="https://mcp-app-store.keenethics-labs.com/app/drivethrough"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>
Markdown
[](https://mcp-app-store.keenethics-labs.com/app/drivethrough)






