Claude / Cowork / MCP Integration with Unraid Shares — Docker Web UI or MCP Server Options?

Hi All

I am looking for some advice on whether anyone in the Unraid community is already doing something similar to this, or whether there are any Docker containers/plugins/projects I should be looking at.

Background

I work in finance and recently started a new role at a smaller business that is very proactive with AI and software adoption.

The company is currently moving from ChatGPT to Claude, mainly because of Claude’s connectors and the way it can interact with business systems. For context, I currently have Claude connected to Slack, OneDrive, Google Workspace (Gmail, Calendar, Google Drive), etc.

From a security/data perspective, my understanding is that Claude Team / Claude for Work data is not used for model training by default, unlike some consumer/default AI products:

https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training

Over the last few weeks, the way I carry out my finance role has changed significantly. For example:

• Document review — I no longer need junior members of the finance team to summarise large documents such as contracts, lease agreements, finance agreements, etc. Claude can produce the first draft and I review/check the output.

• Context gathering — Claude often has more context than a new junior team member, because it can pull relevant information from Gmail, Slack, Google Drive and OneDrive.

• Excel work — I no longer have to build every complex workbook manually. I can tell Claude where the month-end data is, which template to roll forward, and what outputs I want. When working directly in Excel, I can ask it to make changes across multiple tabs while considering the wider workbook context.
  

I have only recently started this role, but I can see that a large amount of data entry, month-end processing and reporting could be automated over the next few months - this isn't AI hype, it genuinely is the case where you work with data sets and routine tasks that computers are better at than humans.

The main finance role left is likely to be business partnering, review, judgement, and being the human interface with the rest of the business.

Important caveat: I do not think this removes the need for qualified finance professionals. In my view, it only works properly if the person prompting/reviewing Claude understands accounting standards, finance systems, controls, reporting and process design.

Personal Unraid Use Case

This brings me to how I would like to use Claude, particularly Claude Cowork, in my personal/home setup.

My Unraid server already contains a lot of structured files and datasets. In effect, these shares are the datasets I would like Claude to interact with. At the moment, Anthropic’s Cowork GUI seems to be aimed at Windows and macOS desktop use. Ideally, I would like something closer to the following:

Unraid server
└── Docker container / Web UI / MCP server
    └── Controlled access to specific shares only
        ├── AI_Workspace
        ├── Project folders
        └── Output folders

The ideal setup would be:

• Claude accessible through a browser-based Web UI.

• Claude able to access only specific Unraid shares, not the whole server.

• A dedicated “working area” share for AI tasks.

• Potential use of local or remote MCP servers so Claude can interact with other services/containers in a controlled way.

• Ideally something containerised, auditable and easier to control than giving a desktop app broad access to mapped network drives.
  

For context, MCP is the Model Context Protocol, which is intended to let AI applications connect to external systems, tools, files and workflows in a standardised way. Claude also supports custom connectors using remote MCP across Claude, Cowork and Claude Desktop, although that feature is still described as beta.

Example Use Cases

Some examples of what I would like to do:

1. Process files dropped into a specific Unraid folder and turn them into a spreadsheet or report.

2. Summarise new documents added to a project folder.

3. Produce a daily digest of Docker container errors/warnings.

4. Check the status of selected Docker apps and provide a plain-English update.

5. Search a defined project share and answer questions using only the files in that share.

6. Potentially interact with other containers through MCP/API wrappers, rather than giving Claude direct shell or Docker socket access.

Security Boundaries Required

I would not want Claude to have access to the whole server. Ideally it would only have access to its own share/project workspace:

/mnt/user/AI_Workspace

Rather than:

/mnt/user/appdata
/mnt/user/system
/mnt/user/domains
/mnt/user/backups
/boot
/var/run/docker.sock

I am also conscious that a generic filesystem tool may be too broad. A safer approach may be a purpose-built MCP server exposing only specific tools such as:

list files
read file
write output file
search workspace
summarise logs
get docker status

Rather than giving it unrestricted shell access.

Questions

Has anyone here implemented anything similar on Unraid?

Specifically:

1. Is there already a Docker container that provides a Claude/Cowork-like Web UI with access to mounted Unraid shares?

2. Has anyone set up a remote MCP server on Unraid for Claude?

3. Are there any Community Apps, plugins or GitHub projects I should be looking at?

4. Is the better approach simply to run Claude Desktop/Cowork inside a Windows VM and map a single Unraid share into that VM?

5. Has anyone safely exposed Docker/container status to an AI assistant without giving it direct Docker socket access?

6. Are there any obvious security pitfalls I should be thinking about before experimenting with this?

At the moment, my fallback idea is:

Windows VM on Unraid
→ Claude Desktop / Cowork
→ Mapped network drive to one dedicated Unraid share only
→ No access to appdata/system/backups

But I would prefer a cleaner Docker/MCP/Web UI approach if one exists.

Interested to hear whether anyone has already gone down this route.

Thanks!