laserfiche-mcp

Community project — not affiliated with or endorsed by Laserfiche.

Let Claude (Desktop, Code, or any MCP client) search and read documents in a self-hosted Laserfiche repository — and, if you opt in, change them. The same program is also a command-line tool (ls, cat, search, dedupe, ...) for work that needs no AI at all.

Read-only by default. Writes are off until you set LF_READ_ONLY=false, and destructive actions always need a second confirmation step. Current release: v2.6.0 — see the CHANGELOG.

Quick start

uv tool install laserfiche-mcp   # or run ad hoc: uvx laserfiche-mcp
laserfiche-mcp setup             # asks for your server + account, then tests the connection
laserfiche-mcp ls 1              # list the repository root

Then connect it to Claude. Prefer no terminal? Use the Claude Desktop extension. New to MCP? Read the step-by-step guide (~7 min). Want ready-made prompts and commands? See examples.

What it can do

Once connected, you can ask Claude things like:

• "Which leases mention an unpaid balance, and what do they say?"
• "Summarize the newest document in the Onboarding folder."
• "Are these two scans the same record, or is something different?"
• "Find duplicate invoices under 2023 and tell me how much space they take."
• "Set Status to Approved on these five entries."

Behind that, it can:

Find things

• Search with native Laserfiche queries or by name pattern — or let Claude draft the query from your repository's own templates and field names, with automatic repair if Laserfiche rejects it.
• Search inside documents, including scans. Full-text search returns the matching passages (page number plus surrounding text) straight from Laserfiche's OCR index, so most "what does it say about X?" questions are answered without downloading anything.
• Look at images. Claude can view picture files in the repository (PNG, JPEG, GIF, WebP) to describe, label and tag them against your own tag and template standard. Images cost tokens, so you're warned of the estimated cost and asked first; for scans of text it tries Laserfiche's OCR before looking at the picture. Optional laserfiche-mcp[images] adds downscaling.

Read things

• Browse folders, look up entries by ID or path, and read metadata, template field values, and your repository's field / tag / template / link definitions.
• Download files, or extract text — PDF, Word, PowerPoint, email, HTML, RTF and plain text (Excel and Outlook .msg need pip install 'laserfiche-mcp[office]') — with page ranges for long documents.

Analyze

• Compare two entries' metadata and fields, and find byte-identical duplicates (only files whose sizes match are ever downloaded). From the command line you can also inventory a whole folder tree to CSV.

Make changes (opt-in)

• Create folders, import documents, copy, rename and move entries.
• Edit fields, tags and links (merge or replace), and assign or clear templates — with required-field checks before Laserfiche sees the request.
• Delete entries, files, or page ranges. Deletes, renames and moves always show a preview first; folders can be fenced, tools allow-listed, and bulk deletes capped. See Safety.

Run where you need it

• Locally in Claude Desktop or Claude Code, as a one-click Desktop extension, or as an HTTP server for claude.ai and ChatGPT with per-user sign-in.
• Self-hosted Laserfiche (Repository API v1 and v2); Cloud support is in beta.
• As a command-line tool with no AI involved — the same operations from a shell, with --json output for scripts. See docs/cli.md.

Every tool, with details: docs/tools.md.

Install

Claude Desktop extension

No terminal, no config files.

1. Download the extension (needs Claude Desktop).

2. Double-click it, click Install, and fill in the form:

   Field	What to enter
   Repository API URL	Your server address, e.g. https://your-server/LFRepositoryAPI
   Repository name	The repository you pick when signing in to Laserfiche Web Access
   Username / Password	An account that can read the repository (the password goes in your OS keychain)

   Not sure what to enter? Ask whoever runs Laserfiche at your organization.

3. Ask Claude something, like "Find every invoice from March in the Accounting folder."

More detail: docs/desktop-extension.md.

Python package

uvx laserfiche-mcp            # run directly, no install
pip install laserfiche-mcp    # or add it to your environment

Needs Python 3.10+ and a reachable Laserfiche Repository API Server with an account that can read it.

Configure

Run laserfiche-mcp setup, or set these environment variables:

Everything else — write guardrails, size limits, search tuning, web-client links, logging — is in docs/configuration.md.

Connect to Claude

Claude Code

claude mcp add laserfiche -- uvx laserfiche-mcp

Pass settings with --env LF_REPO_API_URL=... flags, or export them in your shell first.

Claude Desktop — add this to claude_desktop_config.json (%APPDATA%\Claude\ on Windows, ~/Library/Application Support/Claude/ on macOS), then restart:

{
  "mcpServers": {
    "laserfiche": {
      "command": "uvx",
      "args": ["laserfiche-mcp"],
      "env": {
        "LF_REPO_API_URL": "https://lf.example.com/LFRepositoryAPI",
        "LF_REPOSITORY_ID": "my-repo",
        "LF_API_VERSION": "v1",
        "LF_USERNAME": "service-account",
        "LF_PASSWORD": "replace-me",
        "LF_AUTH_MODE": "password",
        "LF_READ_ONLY": "true"
      }
    }
  }
}

Try a tool without Claude using the MCP Inspector: npx @modelcontextprotocol/inspector uvx laserfiche-mcp

Web clients (claude.ai, ChatGPT)

By default the server talks over stdio to a local client. Web connectors can't launch a local program, so run it as an HTTP server instead:

laserfiche-mcp --http          # serves http://127.0.0.1:8000/mcp

For anything beyond your own machine you need TLS, authentication (per-user OAuth or a shared token) and a network path to your Laserfiche server. Setup and the security checklist: docs/remote-http.md.

Command line

laserfiche-mcp ls '\HR\Leases'                 # list a folder (by path or entry ID)
laserfiche-mcp get 4821 --to ./lease.pdf       # download a document
laserfiche-mcp cat 4821 --pages 4-9            # print its text
laserfiche-mcp search 'unpaid balance'         # full-text search with excerpts
laserfiche-mcp manifest 1 --out inventory.csv  # inventory a folder tree
laserfiche-mcp dedupe 1                        # find byte-identical documents

Every command takes --json. More recipes in examples; the full reference is docs/cli.md.

Safety

• Writes are off by default.
• When on, you can fence them to specific folders (LF_WRITE_PATHS_ALLOW), limit which write tools exist (LF_WRITE_TOOLS_ALLOWED), and cap folder deletes.
• delete, rename and move return a preview and a short-lived token; nothing changes until the call is repeated with that token.
• Start in a sandbox folder: see the recommended write-mode config.

Where your document content goes: Data handling & privacy. All guardrails: docs/safety.md.

Troubleshooting

Start with one command — it tests your login, probes your server, and tells you what's wrong:

laserfiche-mcp diagnose

On Windows, put repository paths in double quotes: laserfiche-mcp list "\HR\Leases". Claude Desktop logs: %APPDATA%\Claude\logs\ (macOS: ~/Library/Logs/Claude/).

When a tool fails it returns a structured error (not a crash) that tells the model whether to retry, ask you, or give up. See docs/error-contract.md.

Documentation

Roadmap

• Cloud verification — needs someone with a Laserfiche Cloud tenant to confirm the auth flow end to end.
• Per-user access (beta) — LF_AUTH_MODE=oauth_passthrough runs Laserfiche calls as each user instead of a shared account; needs confirming against a live LFDS tenant (details).
• Audit logging — a server-side log of every write call.
• v3.0 — remove the old verb-first tool names (get_entry, ...); only laserfiche_{resource}_{verb} remain.
• Later — Workflow trigger tools; stateless-MCP review for multi-instance deployments.

Contributing

Issues and PRs welcome. Run the tests with:

uv sync --extra dev
uv run pytest && uv run ruff check src tests && uv run mypy src

See CONTRIBUTING.md for setup, where help is most wanted, and the opt-in integration tests. This is a community project, not affiliated with or endorsed by Laserfiche.

License

Released under the MIT License. Copyright (c) 2026 Samuel S. Hernandez.