PaperlessMCP

Stop manually organizing your documents. Let AI do it.

You've got a Paperless-ngx instance. You've got hundreds (thousands?) of documents. You *know* you should tag them, set correspondents, organize them properly. But who has time for that?

PaperlessMCP connects your Paperless-ngx to any MCP-compatible AI. Now instead of clicking through the UI, you just ask:

> "Find all my tax documents from 2023"

>

> "Tag these 50 invoices as 'Business Expense' and set the correspondent to 'Acme Corp'"

>

> "Upload this receipt and figure out what it is"

>

> "What documents am I missing from my insurance folder?"

It's Paperless-ngx on LLM steroids. An interface designed *specifically* for AI to manage your documents while you do literally anything else.

What Can AI Do With Your Paperless?

Everything. Full CRUD on every entity type:

43 tools covering:

• Documents — search, upload, download, update, delete, bulk operations, OCR reprocessing
• Tags — full CRUD with colors, matching rules, and hierarchical parents
• Correspondents — track who sends you stuff
• Document Types — classify invoices, receipts, contracts, whatever
• Storage Paths — organize files with smart templates
• Custom Fields — add your own metadata (dates, amounts, URLs, etc.)

All destructive operations require explicit confirmation. Bulk operations default to dry-run mode, so AI can't nuke your archive by accident.

Is PaperlessMCP Right For You?

Yes, if:

• You run Paperless-ngx (self-hosted or cloud)
• You use any AI assistant that speaks MCP (Claude, or anything else supporting the protocol)
• You have a backlog of untagged documents and feel guilty about it
• You'd rather say "organize this" than click 47 buttons
• You want to query your documents in plain English
• You think computers should work for you, not the other way around

No, if:

• You don't use Paperless-ngx (this isn't a general document tool)
• You enjoy manually tagging documents (weirdo, but respect)
• You don't trust AI with your files (fair; destructive operations require confirmation, and bulk operations default to dry-run)

The sweet spot: You've got Paperless running, you've got an MCP-compatible AI, and you want them to be friends.

Getting Started

You'll Need

1. A Paperless-ngx instance with an API token

*(Settings → Django Admin → Tokens → Create one for your user)*

2. An MCP-compatible AI (Claude Desktop, or anything speaking the protocol)

Option 1: Docker (Recommended)

The fastest path from zero to talking to your documents.

docker run -d \
  --name paperless-mcp \
  --restart unless-stopped \
  -e PAPERLESS_BASE_URL=https://your-paperless.example.com \
  -e PAPERLESS_API_TOKEN=your-token-here \
  -p 5000:5000 \
  -v paperless-outbox:/home/mcp/outbox \
  ghcr.io/barryw/paperlessmcp:vX.Y.Z

> Grab the version from the badge above. The release pipeline also publishes `latest`, but pinning a versioned tag gives you a reproducible deployment.

Connect your MCP client to `http://localhost:5000/mcp` and start talking to your documents.

The `paperless-outbox` volume is where `paperless_documents_export_to_outbox` writes exported files. Without it the exports stay inside the container and no other process can reach them — see Sharing the outbox with another MCP server.

Option 2: Claude Desktop

Add to your config file:

{
  "mcpServers": {
    "paperless": {
      "command": "dotnet",
      "args": ["run", "--project", "/path/to/PaperlessMCP/PaperlessMCP", "--", "--stdio"],
      "env": {
        "PAPERLESS_BASE_URL": "https://your-paperless.example.com",
        "PAPERLESS_API_TOKEN": "your-token-here"
      }
    }
  }
}

Restart Claude Desktop. Look for the tools icon — Paperless should be there.

Option 3: Claude Code

One command if you're already running the server somewhere:

# Connect to a running Streamable HTTP server
claude mcp add --transport http paperless http://localhost:5000/mcp

Or run from source with stdio:

claude mcp add --transport stdio paperless \
  -e PAPERLESS_BASE_URL=https://your-paperless.example.com \
  -e PAPERLESS_API_TOKEN=your-token-here \
  -- dotnet run --project /path/to/PaperlessMCP/PaperlessMCP -- --stdio

Verify it's there:

claude mcp list

Option 4: LiteLLM Proxy

LiteLLM can register PaperlessMCP as a Streamable HTTP MCP server in `config.yaml`.

Start PaperlessMCP first using Docker, Kubernetes, or source, then add it to LiteLLM:

mcp_servers:
  paperless:
    url: "http://paperless-mcp:5000/mcp"
    transport: "http"
    description: "Paperless-ngx document management"

Use a URL that the LiteLLM process can reach. In Docker Compose, set the host to the PaperlessMCP service name from that Compose file, such as `paperless-mcp`. If LiteLLM runs directly on the host and PaperlessMCP publishes port 5000, use `http://127.0.0.1:5000/mcp`.

Set `transport: "http"` explicitly for PaperlessMCP's `/mcp` endpoint. LiteLLM's MCP config defaults to `sse`, which is the wrong transport for this endpoint.

`PAPERLESS_API_TOKEN` belongs on the PaperlessMCP service; it is the token PaperlessMCP uses when calling Paperless-ngx. PaperlessMCP does not require an inbound token on `/mcp` unless you put a separate auth layer, such as a reverse proxy, in front of it.

For LiteLLM database-backed MCP storage, enable database storage in LiteLLM:

general_settings:
  store_model_in_db: true

For static configuration, keep the server under the top-level `mcp_servers` key.

Option 5: Kubernetes

For the homelabbers running k8s. We include ready-to-use manifests with Kustomize support.

# Clone and customize
git clone https://github.com/barryw/PaperlessMCP.git
cd PaperlessMCP/k8s

# Customize the checked-in manifests:
# - Set PAPERLESS_BASE_URL in secret.yaml.
# - Pin a versioned image tag in deployment.yaml.
# - The image is public, so remove imagePullSecrets unless your cluster
#   provides the referenced ghcr-secret.

# Create the API token secret (it is not managed by kustomization.yaml)
kubectl create secret generic paperless-token \
  --from-literal=token=your-api-token-here

# Deploy
kubectl apply -k .

See the manifests

Includes: Deployment, Service, Ingress, base-URL Secret, and Kustomization. Tweak to taste.

Option 6: From Source

For contributors and tinkerers:

git clone https://github.com/barryw/PaperlessMCP.git
cd PaperlessMCP
dotnet run --project PaperlessMCP             # Streamable HTTP on :5000
dotnet run --project PaperlessMCP -- --stdio  # stdio mode

Requires .NET 10 SDK.

The Full Toolbox

43 tools, organized by what they touch. Every entity supports full CRUD.

Documents — the main event

Tags — organize everything

Correspondents — who sends you stuff

Document Types — invoices, receipts, contracts...

Storage Paths — where things live

Custom Fields — your own metadata

Health — is it alive?

Configuration

Environment variables. That's it. No config files to manage.

Aliases supported: `PAPERLESS_URL` and `PAPERLESS_TOKEN` also work if that's your style, and `OUTBOX_DIR` is accepted for `PAPERLESS_OUTBOX_DIR`.

Sharing the outbox with another MCP server

`paperless_documents_export_to_outbox` downloads a document server-side and writes it to `PAPERLESS_OUTBOX_DIR`, returning `{path, filename, mime_type, size_bytes}`. The point is that the bytes never travel through the model's context: another tool (a mail server that attaches files by path, say) reads the file directly.

That only works if both containers see the same directory. Mount one volume into both, and make sure the path the other server is told to read matches the path it sees:

services:
  paperless-mcp:
    image: ghcr.io/barryw/paperlessmcp:vX.Y.Z
    environment:
      PAPERLESS_BASE_URL: https://your-paperless.example.com
      PAPERLESS_API_TOKEN: your-token-here
      PAPERLESS_OUTBOX_DIR: /home/mcp/outbox
    ports:
      - "5000:5000"
    volumes:
      - outbox:/home/mcp/outbox

  some-other-mcp:
    image: example/other-mcp:latest
    volumes:
      - outbox:/home/mcp/outbox

volumes:
  outbox:

Two things to know before you rely on it:

• Names carry the document id. A derived name gets the id inserted before the extension (`invoice.pdf` becomes `invoice_42.pdf`), so two documents whose file has the same name cannot overwrite each other. Re-exporting the same document replaces its own file. A `filename` you pass yourself is used as given, so repeated exports under one name do replace each other.
• The archived version is named as such. With `original=false` (the default) Paperless serves the archived PDF, so the export is named after the archived file rather than after a `.jpg` or `.docx` original. Pass `original=true` to get the uploaded file under its own name.
• Exports appear whole. The download is streamed to a temporary file in the outbox and renamed into place, so a reader on the other side of the volume never picks up a half-written file, and a symlink planted at the destination is replaced rather than written through.
• The directory must be writable by the container user. The image runs as root unless you override it, so exports land in a bind mount owned by root — if the consuming container runs as a non-root user, set `PAPERLESS_OUTBOX_DIR` to a directory both can write, or fix the ownership yourself. The directory is created on first export, and a failure surfaces there rather than at startup.

LocalAI Compatibility

Streamable HTTP clients are expected to send `Accept: application/json, text/event-stream` on `/mcp` POST requests. Some clients cannot configure that header. Set `MCP_RELAX_ACCEPT_HEADER=true` to have PaperlessMCP normalize missing or incomplete `Accept` headers before the MCP SDK handles the request.

Support the Project

If PaperlessMCP saves you time, consider supporting development:

Every bit helps keep the lights on and the commits flowing.

Contributing

Yes please. We use trunk-based development with conventional commits.

git clone https://github.com/barryw/PaperlessMCP.git
cd PaperlessMCP
dotnet build
dotnet test

The rules:

• Conventional commits (`feat:`, `fix:`, `docs:`, etc.) — versions bump automatically
• Tests pass or it doesn't merge
• Destructive operations need `confirm=true`; bulk operations default to dry-run

See CONTRIBUTING.md for the full rundown.

License

MIT — do whatever you want, just don't blame me.

Acknowledgments

• Paperless-ngx — the document system that makes this worth building
• Model Context Protocol — the glue between AI and everything else
• Everyone who's ever felt guilty about their untagged documents