Official Nextcloud MCP Server

[andrea-tr2]

Is your feature request related to a problem? Please describe.
The Model Context Protocol (MCP) is becoming the de-facto standard for letting LLMs query real-world data.

Nextcloud Assistant’s Context Chat already answers questions about files kept inside Hub, but any external LLM must fall back to bespoke REST/WebDAV wrappers. Each team re-implements connectors, duplicates ACL logic, and risks leaking or corrupting data. 

Describe the solution you’d like
An official @nextcloud/mcp-server that:

• speaks MCP through read-optimised endpoints only;
• honours existing share & group permissions for files, Talk messages, etc.;
• installs with a single docker run ... or npx ... and token-based auth;
• includes starter configs for LangChain, Claude and OpenAI Assistants so devs can plug-in immediately.

Describe alternatives you’ve considered

• Keep using raw REST/WebDAV plus custom ACL plumbing (high maintenance).
• Expose Context Chat’s internal API (helps only in-app queries).
• Rely on community wrappers (fragmented, inconsistent).

[ChristophWurst]

cc @DaphneMuller

[DaphneMuller]

hi, thank you for reaching out. Some discussions already took place to use this for the Context Agent feature, feel welcome to continue the discussion there (https://github.com/nextcloud/context_agent) or in the general Nextcloud Assistant repository (
https://github.com/nextcloud/assistant/ )

[andrea-tr2]

Hi @DaphneMuller , thanks for pointing that out!

The confusion seems to come from mixing up Context Agent/Assistant with the Model Context Protocol (MCP), but they solve very different problems:

• Nextcloud Context Agent + Assistant are in-Nextcloud apps. They let the built-in Assistant trigger actions by calling internal Nextcloud APIs. They only work once you’re already inside a Nextcloud session and rely on the Assistant UI workflow.  
• MCP is an open, vendor-neutral wire-protocol that any external LLM client (LangChain, Claude, OpenAI Assistants, etc.) can speak to directly.

Because MCP lives outside the Nextcloud PHP stack, it can:

1. Reuse existing ACLs instead of re-implementing them in every plugin.
2. Offer a token-protected gateway that is safe for third-party LLMs.
3. Let any LLM-powered product connect with a single docker run nextcloud/mcp-server ..., instead of writing a new WebDAV/RPC adapter.

In short, Context Agent is about inside-Nextcloud automation, while MCP is about standardised, external context retrieval.

Does it make sense?

[DaphneMuller]

With context agent we also retrieve some external sources like the weather, a duckduckgo search, or youtube. If you think it belongs here feel welcome to make a more concrete proposal for how this would work in the code (if you're up for doing the development work)

[andrea-tr2]

Hey @DaphneMuller, I think we’re still talking about two completely different layers :)

• Nextcloud Context Agent is an in-app action layer. It lets the in-Nextcloud Assistant hit weather, DuckDuckGo, YouTube, etc., then push the results back into the chat or fire off internal actions. It’s built as a Nextcloud app and speaks private PHP/OCS APIs.
• MCP (the widely adopted Anthropic-led open standard) is a wire-protocol for external LLMs. It standardises a "USB-C port" so tools like LangChain, Claude or OpenAI Assistants can pull files/Talk threads directly, re-using Nextcloud’s shares and ACLs, without everyone writing their own WebDAV+ACL wrapper.

So... Agent = actions inside Nextcloud. MCP = context outside Nextcloud. They’re complementary, not overlapping.

For reference, check the fresh Directus implementation: https://github.com/directus/mcp

For proof this is already catching on, there are two community/non-official MCP bridges for Nextcloud:

• https://github.com/cbcoutinho/nextcloud-mcp-server
• https://github.com/Monadical-SAS/nextcloud-mcp-server

[oleksandr-nc]

As far as I understand this is a request for a new Nextcloud application(App or ExApp) that can expose required and relative information for a user to an AI using MCP.

Am I understanding correctly @andreatorre-dev ?

[andrea-tr2]

Hi @oleksandr-nc thanks for the follow-up! Just to be clear: no, this isn't another PHP app inside Nextcloud. It's a lightweight sidecar you'd launch with something like docker run nextcloud/mcp-server ..., sitting next to your instance and speaking the new Model Context Protocol (MCP).

Quick context. Over the past few months:

1. Anthropic introduced MCP — a universal "USB-C" plug for AI to read data.
2. OpenAI adopted it right away; it's built into their Agents SDK and the ChatGPT desktop app.
3. Google followed with Agent-to-Agent (A2A) and says A2A works alongside MCP: A2A handles agent co-ordination, MCP moves the data.

So... what nextcloud-mcp-server (Streamable HTTP transport) would actually do is relatively simple: it checks an OpenID Connect token to see who's calling, pulls the requested file (or Talk thread, etc.) through the existing WebDAV/OCS endpoints (so your shares and permissions stay exactly as they are) and hands the result back in MCP's minimal JSON format.

Because it relies entirely on Nextcloud's current APIs and ACLs, nothing in core or in Context Agent needs to change. The bridge just lets any MCP-ready LLM plug in without custom WebDAV plumbing.

Hope that clears it up!

@cbcoutinho, @a9tavako please feel free to clarify further.

[cbcoutinho]

Thanks for opening this feature request and the mention of nextcloud-mcp-server! I appreciate the opportunity to share insights from building and running this for myself.

I agree with @andreatorre-dev that MCP is becoming the de facto standard for LLM integration. What drew me to implement this was a specific workflow need—I wanted my AI-powered development environment (VS Code with Cline extension) to maintain persistent insights across projects and investigations. The Nextcloud Notes App and MCP protocol turned out to be a perfect fit for this use case, though I've learned quite a bit about its current limitations along the way.

How I've Deployed It
Currently, I'm running both Nextcloud and the MCP server in Kubernetes using the SSE/HTTP transport. It's been working well for my single-user setup, though you could just as easily run it locally via stdio and connect to any Nextcloud instance. Right now it uses Basic Auth with environment variables, which keeps things simple when deploying via Docker but obviously limits it to one user.
I'd love to implement OAuth2 for multi-user support, but honestly, the VS Code extension I use doesn't support it yet, so there's no immediate pressure. When that changes, it'll be a natural next step.

Real-World Challenges I've Hit
Building this has taught me a lot about both MCP and Nextcloud's API landscape:
The biggest hurdle has been authentication inconsistencies. My server primarily uses the Notes App REST API, but sometimes needs to hit WebDAV directly. The problem? Notes API doesn't support Authorization: Bearer tokens while WebDAV does. This makes OAuth2 implementation trickier than it should be—it seems like OAuth2 support isn't consistently available across all of Nextcloud's APIs.
On the MCP side, while the protocol is elegant for basic use cases, production concerns like session handling, authentication patterns, and resource management aren't as well-defined as I'd like. The ecosystem is moving fast, but there are still rough edges.

What I'd Love to See in an Official Version
An official Nextcloud MCP server would solve a lot of these pain points:

1. Consistent OAuth2 support across all APIs would make multi-user scenarios much cleaner
2. Read-optimized endpoints (as you mentioned) would definitely improve performance—some of my current API calls are heavier than they need to be
3. Native permission integration rather than having to navigate different ACL models

Hope this helps!

[leomixe]

@DaphneMuller: With a fully integrated MCP server - for all Nextcloud-Apps (as Notes, Deck, Talk...) it would behave as follows from the view of a user:
ChatGPT (or the LLM of you choice) is just a additional user of Nextcloud e.g. I would call it ChatGptBot. You can e.g. assign a Deck board to the user ChatGptBot. Then you log into your ChatGPT-Account and add a connector to your MCP-Nextcloud Server. Now your ChatGpt-Account is connected to Nextcloud. From within ChatGpt you can just talk "please be my project manager and configure my deck board with some new cards for my new project...."
Using this method you have the full control which files, decks or notes are visible to the ChatBot. You have to share it.
MCP server is no longer a read only solution it has been expanded also to interact and write! It will behave like an intelligent remote control of Nextcloud. The read/write access to Nextcloud notes is already implemented into https://github.com/cbcoutinho/nextcloud-mcp-server
A good start point to implement MCP-interfaces to other Nextcloud-Apps.

[andrea-tr2]

For context, ChatGPT workspace admins can now (June 2025) add team-wide MCP servers.

https://platform.openai.com/docs/mcp