MCP in depth

The tools an MCP client gets, what each permission allows, how Tandmade checks every write, and the limits to know before you build a workflow on it.

This continues Use your own AI agent. The server speaks MCP over streamable HTTP at https://app.tandmade.com/mcp, with OAuth sign-in through your Tandmade account.

Permissions

Permission Default Allows
Read always sites, structure, code, design system, brand, analytics, history
Content on collection items and media
Comments on reading, writing, answering, resolving
Code on changing files, sections, tokens, uploading files
Publish off a collection item live; the site to its review link or live site

A Content editor gets read, content and comments only. A site handed over to a client without a plan isn't served.

Tools

About 40 tools, grouped:

Area What the tools do
Overview list sites; the structure of pages, sections and components; the site's code rules; the design system and what is off-system
Reading code read a file or its outline, search the code, check a file against the design system
In a browser render a page and find errors; screenshot a page or element, also in hover and on mobile; sizes and computed styles
Changes edit files as one history step; add a section; set a section's property; reorder sections; edit design tokens; undo your own step
Files add images, icons, video and files (up to 30 files and 60 MB in one step); add a custom font
Content collections, items and media; create (always a draft), edit and delete items; upload a photo or PDF
Comments read, comment on an element with a mention, reply, resolve and reopen
Analytics traffic with sources, pages, devices, AI crawlers, 404s and the effect of each publish; one page in detail
History every change by people and apps
Brand read the brand part by part; add a rule in the user's words ("We never say cheap"), with undo
Publishing an item live; the site to the review link (default) or the live site

How writes are checked

  • The design system is enforced on the server. It refuses a hard-coded color, a variable that doesn't exist, a font by name, a broken import, or deleting a file that's still in use, and says why. A value that equals a token is turned into the token.
  • After a write, up to five affected pages are rendered before and after. A page that breaks rolls the step back by itself.
  • Every call is a history step with its author and the app's name. People in the editor see the app's face in the top bar and its steps on the page, and their own undo never takes back the app's step.
  • Changes land in the working copy, as in the editor. Content is saved as drafts.
  • Publishing needs the permission and an explicit request. The default target is the review link, and the client asks before anything goes out.
  • Secrets and form submissions are never readable.

Good practice

  • Ask the app to read the site's structure and code rules first; the server's answers then follow your conventions.
  • Keep one connection per workspace you work in.
  • Use the editor to review: open the site, look at the app's changes on the page and undo what you don't want.

Limits

  • Translations of collection items, item versions and form submissions aren't available as tools.
  • Video files over 25 MB can't be uploaded.
  • Apps without a terminal (claude.ai, ChatGPT) upload files only from a public address, not from your disk.
Did this answer your question? Thanks for the feedback There was a problem submitting your feedback. Please try again later.