Nextcloud MCP server

What is an MCP server for Nextcloud?

An MCP server for Nextcloud is a lean bridge between your AI model and your Nextcloud instance. It translates the agent's requests into standardised calls against Nextcloud's WebDAV and OCS API, the two interfaces that the official desktop and mobile clients also use. Instead of maintaining separate scripts for each AI integration, all models use the same clearly defined tools through the open Model Context Protocol.

The decisive point with Nextcloud is not primarily the functionality but the location: Nextcloud is self-hosted collaboration. Files, Talk and groupware (calendar, contacts, notes, Deck) run on your own server or with an EU host. Anyone who runs Nextcloud has deliberately decided against Microsoft 365 or Google Workspace. An MCP server has to respect that stance, and it does, because it simply runs alongside in the same data centre.

Use cases: what can a Nextcloud MCP server do in practice?

1. Document search across the whole file storage “Where is the signed contract with supplier X?” The agent walks through the folder structure via PROPFIND, finds the right file, fetches it via WebDAV GET and summarises the relevant section, instead of someone clicking through many subfolders.
2. File workflows & storage hygiene Incoming documents are recognised, renamed according to convention and moved to the right project folder via MOVE. The agent reports duplicates and orphaned folders, so the storage stays tidy without a manual routine.
3. Notes & knowledge management From meeting minutes, the agent creates structured notes in the Nextcloud Notes app, assigns tags and links the related files. Internal knowledge grows without anyone having to maintain it afterwards.
4. Public share links on request “Share the quarterly report with the tax advisor, link valid for seven days.” Through the OCS Shares API, the agent creates a password-protected, expiring share link, controlled and logged.
5. Talk summaries & status queries The agent summarises long discussions in Nextcloud Talk, pulls out open to-dos and answers “What was decided most recently in the project channel?”, as a read tool, without posting itself.
6. Deck boards & task upkeep From an email or note, the agent creates a card on the appropriate Deck board and sets the due date and labels. Tasks land where the team already works, without switching tools.

How the integration works

Architecture overview for Nextcloud:

AI agent ↔ MCP server (your EU / on-prem host) ↔ Nextcloud (WebDAV at /remote.php/dav + OCS API)

Concrete technical steps:

• App password instead of main login: We create a dedicated user for the MCP server in Nextcloud, grant it access to only the necessary folders and apps and use its app password, which can be revoked individually at any time. Where your instance offers OAuth2/OIDC, we use that as an even cleaner variant.
• Tool mapping: We define MCP tools such as nc_webdav_list, nc_read_file, nc_webdav_move, nc_create_note and nc_create_share. Each with a JSON schema, validated inputs and clear error messages.
• WebDAV for files: Listing via PROPFIND, reading via GET, writing via PUT, moving via MOVE: exactly the path the Nextcloud clients take. Nextcloud's own properties in the oc: and nc: namespaces (tags, favourites) can be used as well.
• OCS for the rest: Shares, Notes, Deck and Talk run through the OCS API. Read and write access is enabled per app, not as a blanket permission.
• Read vs. write separation: By default, the agent reads and drafts. Deleting, moving large volumes or public shares go through an approval level, or are disabled entirely.
• Logging: Every API action is logged with timestamp, user and affected path.

GDPR, data residency and security

This is where Nextcloud plays to its real strength: the data is already where you want it. An MCP server must not weaken this advantage through a back door. We watch for this in every Nextcloud integration:

• Data residency up to the MCP server: Nextcloud stays on-premise or with your EU host, and the MCP server sits in the same data centre. No US edge between Nextcloud and the MCP server.
• Restricted identity: The agent works with an app password or OIDC token with minimal permissions, never with the admin or user password.
• AI model: this is where data residency is decided: If you want no US cloud in the path, run a local open-weight model. Alternatively, we use Claude via AWS Bedrock (EU region Frankfurt), Google Gemini via Vertex AI (EU), OpenAI Enterprise EU or Azure OpenAI in an EU region, with a DPA (data processing agreement). The requests then run in EU data centres, but with a US provider.
• Data minimisation: The agent sees only the folders and apps needed for the task, with no full access to the whole instance by default.
• Client / group separation: With several departments or clients in one Nextcloud, access can be cleanly delimited per group.

For many organisations that run Nextcloud precisely for sovereignty reasons (law and tax firms, public-sector-related bodies, research institutions, associations), this is often the only acceptable way to let AI near their documents at all.

What does it cost?

Nextcloud can be integrated well through its documented standard APIs. The effort depends mainly on how many apps (files only, or also Talk/Notes/Deck) and how many write workflows you need. In concrete terms:

• First use case clearly scoped, then gradual expansion: typically you start with document search and file workflows, because they deliver benefit fastest.
• Clear scope per use case, no hours roulette. Scaling after that in small, clearly defined steps.

A concrete figure is part of the initial consultation. It depends on the scope of apps, the number of write tools and whether the AI model is to be hosted or run locally.

Frequently asked questions

Is there an official Nextcloud MCP server?

Yes. According to the Nextcloud administration manual, the Nextcloud Context Agent app provides an MCP server for external language models and applications. It runs as an external app via AppAPI, is reachable at /index.php/apps/app_api/proxy/context_agent/mcp/ on your instance and requires an app password as a bearer token. There are also open-source MCP servers from the community. We build our own MCP server against the official WebDAV and OCS API if you want to define the tools, permissions and audit log yourself.

Which Nextcloud interface is used?

For files, the WebDAV API at /remote.php/dav/files/{user}/; for shares, notes, Deck and Talk, the OCS API. Both are present in every standard Nextcloud and documented, and no special plugin is needed.

Does the data really stay in Germany?

Yes, if the AI model runs there too. Nextcloud runs on-premise or with an EU host of your choice, and the MCP server sits next to it. With a locally run model, nothing leaves your own premises. If you use Claude via AWS Bedrock, Gemini via Vertex AI or Azure OpenAI instead, a US provider processes the requests in its EU region; the data centres are then in the EU, but a US provider is in the path.

How does the agent authenticate?

Via an app password of a dedicated Nextcloud user for the agent, whose permissions are limited to the necessary folders and apps. You can revoke the app password individually at any time. Where your instance offers it, we alternatively use OAuth2/OIDC. The normal user or admin password is never passed on.

Can the agent also delete or move files?

Technically yes, but in practice we strictly separate read and write permissions. Destructive operations such as deleting or mass moves go through an approval level or are disabled completely. The default is “read and draft”.

Which AI models work?

All MCP-compatible models: Anthropic Claude, OpenAI GPT, Google Gemini, Microsoft Copilot, and, in keeping with the sovereignty logic, local open-weight models too. A change of model is possible without a new integration.