Plone MCP (Model Context Protocol) Server

[danalvrz]

Important

If you are not a member of the Volto Team or Developers Team in the Plone GitHub organization, then do not work on or comment on this issue.

PLIP (Plone Improvement Proposal)

Responsible Persons

Proposer: Dante Álvarez (kitconcept GmbH), @danalvrz

Seconder: Timo Stollenwerk (kitconcept GmbH), @tisto

Abstract

This proposal outlines the implementation of a Model-Context Protocol (MCP) server as a core feature for Plone. This server acts as a standardized interface, enabling Large Language Models (LLMs) and other AI agents to interact with a Plone site in a structured way. It exposes Plone's capabilities as a set of tools to the AI world.

Motivation

By adopting this MCP server, we allow users things such as:

• Conversational Content Creation/Editing: Users can instruct an AI assistant in natural language.
• Automated Content Auditing & Management.

Assumptions

This implementation relies on the functionality of plone.restapi, as the MCP server uses it as its underlying connection to the backend.

Proposal & Implementation

We propose the creation of a new package plone-mcp, which will contain and manage the MCP server application.

Deliverables

• A new, distributable plone-mcp package.
• Documentation.

MVP (Minimal Viable Product)

Basic authentication, adding content objects, adding blocks, and basic workflow transitions:

• As a user, I can add the Plone MCP to my local Desktop AI client (Claude Code, ChatGPT)
• As an anonymous user, I can authenticate to an existing Plone instance with username and password
• As a logged-in user, I can add a Page to a Plone site
• As a logged-in user, I can add a slate block to a page
• As a logged-in user, I can add an image block to a page
• As a logged-in user, I can add a teaser block to a page
• As a logged-in user, I can add a grid block to a page
• As a logged-in user, I can publish a private page
• As a logged-in user, I can retract a published page

Later

• As a logged-in user, I can add a News Item to a Plone site
• As a logged-in user, I can add an Event to a Plone site
• As a logged-in user, I can add an Image to a Plone site
• As a logged-in user, I can add a File to a Plone site
• As a logged-in user, I can add a Redirect to a Plone site
• As a logged-in user, I can add a listing block to a page
• As a logged-in user, I can add a search block to a page
• As a logged-in user, I can add a table block to a page
• As a logged-in user, I can add a video block to a page
• As a logged-in user, I can add a maps block to a page

Risks

• Security: As a powerful new entry point into the system, the MCP server endpoint must be rigorously secured.
• Performance: High-frequency requests from an AI agent could place a significant load on Plone sites.

Participants

@tisto
@danalvrz
@rohnsha0
@Tishasoumya-02

Decisions (that we need to make)

• Language: Do we want to use JS or Python for the implementation of the MCP server?
• Naming: plone-mcp, plone-mcpserver, or something different?
• MVP feature set: tbd
• How do we distribute the MCP server?
  • Repo checkout, building, and running it locally
  • Docker Container running
  • Cloudflare or similar URL -> talk to the AI team about it

[davisagli]

I'll move this to the CMFPlone repository, since it isn't a PLIP about Volto.

[davisagli]

This is fine as a starting point for conversation but I'd like to see some more detail in the proposal:

• What would you consider as the minimum viable set of functionality to be exposed via the MCP server?
• Some examples of prompts that you want to support would be helpful for communicating about the feature as well as provide a starting point for test cases.
• How do you plan to mitigate the risks?
• What language/tools will you use for implementing the MCP server?

We propose the creation of a new Plone package, plone.mcp.server

I don't think there's a reason to create a new namespace (plone.mcp). It can be plone.mcpserver or plone.mcp_server or plone-mcpserver (assuming it's a Python package and not something else).

A new, distributable Plone package.

This sounds like it would be a new command that you can install and run alongside the Plone backend, rather than something which is built in to Plone, right?

Security: As a powerful new entry point into the system, the MCP server endpoint must be rigorously secured.

Does that require any new functionality in the framework, or is it just a matter of making sure that normal permission checks are also applied when the REST API is being called via the MCP server? Are there any operations which should be allowed when carried out by a real person but not when carried out via the MCP server?

Performance: High-frequency requests from an AI agent could place a significant load.

This is also already true of high-frequency requests from a non-AI agent.

[rohnsha0]

@danalvrz Please add me in too!
Having created various MCPs, I love to help and make this happen!
Bt one thing, in addition to @davisagli's queries, I would like to know the transfer protocol you plan to use for the implementation. Since security is to be kept in mind.

[tisto]

This is fine as a starting point for conversation but I'd like to see some more detail in the proposal:

• What would you consider as the minimum viable set of functionality to be exposed via the MCP server?
• Some examples of prompts that you want to support would be helpful for communicating about the feature as well as provide a starting point for test cases.

I started to write down some use cases here:

plone/plone-mcp#8

• How do you plan to mitigate the risks?
• What language/tools will you use for implementing the MCP server?

We have to make a decision if we want to build the MCP server in Python or in JS. For our prototype we choose JS. However, this is open for debate.

We propose the creation of a new Plone package, plone.mcp.server

I don't think there's a reason to create a new namespace (plone.mcp). It can be plone.mcpserver or plone.mcp_server or plone-mcpserver (assuming it's a Python package and not something else).

+1

We have a prototype https://github.com/kitconcept/plone-mcp

Most MCP servers are either named <PROJECT_NAME>-mcp or <PROJECT_NAME>-mcpserver.

A new, distributable Plone package.

This sounds like it would be a new command that you can install and run alongside the Plone backend, rather than something which is built in to Plone, right?

Correct. An MCP server is an extension/add-on for an AI/LLM and installed e.g. in Claude.

Security: As a powerful new entry point into the system, the MCP server endpoint must be rigorously secured.

Does that require any new functionality in the framework, or is it just a matter of making sure that normal permission checks are also applied when the REST API is being called via the MCP server? Are there any operations which should be allowed when carried out by a real person but not when carried out via the MCP server?

I don't see what we need to do beyond reusing Plone's security model. However, this is something we need to investigate.

Performance: High-frequency requests from an AI agent could place a significant load.

This is also already true of high-frequency requests from a non-AI agent.

+1

[tisto]

@plone/framework-team @plone/release-team, we are working on an MCP (Model Context Protocol) server for Plone. See https://modelcontextprotocol.io/docs/getting-started/intro for an introduction. Right now, we are experimenting in a kitconcept repo here:

https://github.com/kitconcept/plone-mcp

To collaborate with the community, we would like to move the repo to the Plone org. Therefore, I would like a vote on the following questions:

• Do we want an official MCP server for the Plone community?
• Do we want to move the current kitconcept implementation to the Plone org on GitHub to allow further development on this?

@davisagli I specified an MVP with user stories for:

• Basic authentication, adding content objects, adding blocks, and basic workflow transitions (see description).

@rohnsha0 I added you to the PLIP description. Looking forward to working on this with you and hearing your thoughts!

[erral]

I would also add @plone/classicui-team to the conversation in case they haven't seen the issue yet.

[tisto]

@erral nothing wrong with mentioning anyone for an FYI. However, didn't we say that the Volto and the ClassicUI team are only involved when a PLIP covers their specific areas? I was under the assumption that the new "virtual" @plone/framework-team would make this decision in a second step, if needed.

Since this is neither Volto nor ClassicUI related, I deliberately did not mention those two teams.

I am just trying to figure out the right process here. Did we ever document our agreement somewhere?

There might be a misunderstanding, or we did not make this clear enough in the description. The plone-mcp package is not a Plone package that will be shipped with Plone in any way. It is a standalone package that can be added to a local desktop AI (e.g., Claude Desktop or ChatGPT Desktop) to connect to any Plone instance.

[erral]

@erral nothing wrong with mentioning anyone for an FYI. However, didn't we say that the Volto and the ClassicUI team are only involved when a PLIP covers their specific areas? I was under the assumption that the new "virtual" @plone/framework-team would make this decision in a second step, if needed.

Since this is neither Volto nor ClassicUI related, I deliberately did not mention those two teams.

I am just trying to figure out the right process here. Did we ever document our agreement somewhere?

There might be a misunderstanding, or we did not make this clear enough in the description. The plone-mcp package is not a Plone package that will be shipped with Plone in any way. It is a standalone package that can be added to a local desktop AI (e.g., Claude Desktop or ChatGPT Desktop) to connect to any Plone instance.

Yes, sorry, you are right, I misunderstood the mcp point here.

I thought as a feature to provide by Plone where also more people could be interested. Sorry for the noise.

[stevepiercy]

I am just trying to figure out the right process here. Did we ever document our agreement somewhere?

You rang? 😁

Current published docs:

• https://6.docs.plone.org/contributing/core/plips.html
• https://6.docs.plone.org/contributing/core/plip-review.html

Stalled issue:

• PLIP Process for User Interfacing Changing PLIPs volto#6628

From the Plone Community Steering Circle - 2025-02-06:

PLIP Process for User Interfacing Changing PLIPs, includes revising the overarching Plone Improvement Proposals (PLIPs) process. There is no more Framework Team.

We even created a Discord channel to complete and publish the process, called plip-process-team, but that also stalled.