Overrides and MCP Tower prompts

Admins and Editors can fix what agents read about a server, and add prompts to it, without waiting for its developers. MCP Tower’s middleware applies the changes to what every agent receives. Your server keeps its own text: scans read it directly, so Inspect always knows both.

What you can change

Tools: the title, the description, and the description of any argument, including nested ones (filters.price.max) up to three levels below the arguments.
Your server’s prompts: the title, the description and the description of each argument.
Resources and resource templates: the title and the description.

Names, schemas, types, required arguments, defaults, hints and _meta are never changed: your server checks calls against its own schema, and hosts guard tools by their hints.

Change a text

Open the server in Inspect and expand the tool, prompt or resource.
Select the pencil next to a text, or click the text. The editor shows what agents see now, your server’s text under it, and how many characters are left.
Press Enter (titles) or Ctrl+Enter / ⌘+Enter (descriptions) to save, or Escape to cancel.

Saving your server’s own text again removes the override. An empty description is allowed: it hides a description that misleads agents.

Every changed text carries an Overridden mark, on the row (“2 overrides”) and next to the field. Show original puts your server’s text and what agents see side by side, and Reset to original removes the override. Viewers see the same marks without the controls.

After a save the field says “Applying…” until the middleware is published, then “Live”. Agents see the change the next time they load the list, which for most hosts is a new conversation.

Limits: 200 characters for a title, 8,000 for a tool or prompt description, 2,000 for an argument or resource description, and 128 KB of overrides and MCP Tower prompts per server.

When your server changes the original

If a scan finds that your server now sends a different text for a field you overrode, Inspect shows “The server changed the original since this override” with both versions. A tool with such a field counts as needing review. Choose Keep override to keep your text, or Use the new original to remove the override.

An override of a tool, argument or prompt your server no longer offers is kept, marked “Not applied”, and applies again if it comes back.

MCP Tower prompts

Under Prompts, New prompt adds a prompt that the middleware serves next to your server’s:

A name (lowercase letters, digits, _ and -), a title and a description.
Arguments, each with a description and whether it is required.
Messages, as the user or the assistant. Write {{argument}} where a value goes; typing {{ suggests the argument names. Any other {{ is refused when you save.

The preview fills in sample values the way the middleware does. MCP Tower prompts are listed on the last page of prompts/list and answered by the middleware itself; your server never receives their prompts/get calls. A missing required argument answers error -32602. If your server declares no prompts, the middleware declares them for you.

A prompt named like one of your server’s replaces it: agents get yours, and Inspect marks both.

Up to 50 MCP Tower prompts per server, with up to 16 KB of message text each.

Catalog listings

If your server is listed in ChatGPT, Claude or another catalog, an override changes what that catalog reviewed: the live server no longer sends the texts it approved. Check the catalog’s rules for changed descriptions; you may need to resubmit. MCP Tower shows this note the first time you save a change on a live server in a session.

History

Every change is in your organization’s audit log (Settings, Audit log) under Tools, with who made it and the text before and after.