You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
MCP content write tools: let agents ask for a short result instead of the whole entry
#4547
content_create, content_update and content_publish return the complete entry after a write. That includes every field, the draft and live data, and expanded bylines, serialised with JSON.stringify(data, null, 2). content_get does the same, which makes sense for a read. For a write, though, the agent usually only needs to know that the write succeeded and what the new _rev is.
For an ordinary long-form article the result is 55–110 KB. That is more than AI clients will put into the conversation. Claude Code, for example, reports that the output exceeds its limit and saves it to a file. So after a successful write, the agent has to read a file to find _rev before the next update, or it treats the oversized result as a failure. Every write also spends a large part of the context window on content the agent just sent.
Use case
emdashnow.com is a publication about emDash, and we edit it partly through emDash's MCP server with Claude Code. In one session on 30 October 2026, for example, we staged edits to six published articles and created three drafts through content_update and content_create. Every one of those writes returned an oversized result, even though each one succeeded.
This affects anyone using the MCP server for editorial work on real-length content. Short demo posts don't show it.
Current workarounds
Let the client save the result to a file, then search that file for _rev.
Call content_get with a narrower purpose afterwards. That doesn't help, because it returns the same full entry.
Use the REST API from a script instead of MCP for bulk edits. That gives up the MCP server's permission model and tooling.
Possible directions, as examples rather than a design:
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Problem
content_create,content_updateandcontent_publishreturn the complete entry after a write. That includes every field, the draft and live data, and expanded bylines, serialised withJSON.stringify(data, null, 2).content_getdoes the same, which makes sense for a read. For a write, though, the agent usually only needs to know that the write succeeded and what the new_revis.For an ordinary long-form article the result is 55–110 KB. That is more than AI clients will put into the conversation. Claude Code, for example, reports that the output exceeds its limit and saves it to a file. So after a successful write, the agent has to read a file to find
_revbefore the next update, or it treats the oversized result as a failure. Every write also spends a large part of the context window on content the agent just sent.Use case
emdashnow.com is a publication about emDash, and we edit it partly through emDash's MCP server with Claude Code. In one session on 30 October 2026, for example, we staged edits to six published articles and created three drafts through
content_updateandcontent_create. Every one of those writes returned an oversized result, even though each one succeeded.This affects anyone using the MCP server for editorial work on real-length content. Short demo posts don't show it.
Current workarounds
_rev.content_getwith a narrower purpose afterwards. That doesn't help, because it returns the same full entry.Possible directions, as examples rather than a design:
result: "summary", that returns onlyid,slug,status,_rev, and the draft and live revision IDs. That would also be a natural place for the public URL that MCP tools should return the item's public URL (content_get / content_list / content_publish) #2554 asks for.I drafted this with Claude Code (Claude Opus 5.5), then reviewed and edited it myself.
Your involvement
All reactions