laserfiche-mcp

Community project — not affiliated with or endorsed by Laserfiche.

A Model Context Protocol server that lets Claude (Desktop, Code, or any MCP client) search, read, and — when you opt in — write documents in a self-hosted Laserfiche repository. The same binary is also a full command-line client (ls, get, cat, search, manifest, dedupe, ...) for the deterministic work that needs no model at all.

Current release v2.3.0 — read and write tools for self-hosted Repository API v1 and v2, a one-click Claude Desktop extension, an optional remote HTTP transport with per-user OAuth for web clients, and a full CLI (ls, get, cat, search, manifest, dedupe, ...) for the deterministic work that needs no model at all. Read-only by default; write tools register only with LF_READ_ONLY=false and are guarded by path fences and a two-step, parameter-bound confirmation flow. See the CHANGELOG for per-release notes and the roadmap for what's next.

Quick start

uv tool install laserfiche-mcp   # or run ad hoc: uvx laserfiche-mcp
laserfiche-mcp setup             # wizard: server URL + account, then verifies the connection
laserfiche-mcp ls 1              # you're in — list the repository root

Then wire it into your MCP client — see Claude Desktop and Claude Code below. Prefer environment variables over the wizard? See Configure. Want a no-terminal install instead? See the Claude Desktop extension.

What you can do with it

Once connected, Claude can:

Read (always available):

• Search the repository with native Laserfiche search syntax, by name pattern, or via the LLM-friendly search_natural flow (asks the server for templates first, then runs with automatic 400 repair)
• List the contents of any folder, look up an entry by ID or path, read all template field values, list field/tag/template/link definitions and audit reasons
• Inspect document metadata, fetch the raw edoc as base64, or extract text locally (PDF, DOCX, PPTX, XLSX, EML, HTML, RTF, text/*) — all via get_document_edoc(..., mode=...)

Write (opt-in via LF_READ_ONLY=false):

• Create folders, import documents, copy entries (async), rename and move entries
• Set, merge, and clear fields, tags, and links on an entry
• Assign and remove templates — with optional client-side validation of repository-required fields before the API call
• Delete entries (folders cascade), edocs, and specific page ranges — all with a two-step preview→confirm-token flow, HMAC-signed and bound to operation + entry + the operation's own parameters, expiring after 5 minutes

Operate safely — every write checks the entry's path against LF_WRITE_PATHS_ALLOW / LF_WRITE_PATHS_DENY, folder deletes refuse unless force_large_delete=true when child count exceeds LF_DELETE_FOLDER_MAX_DESCENDANTS, and LF_WRITE_TOOLS_ALLOWED can scope a deployment to e.g. metadata-only writes.

Install

Two ways to run it, depending on who you are.

For everyone — the Claude Desktop extension

Chat with your Laserfiche repository from Claude Desktop — no terminal, no config files.

1. Download

Download the extension (always the newest version), or browse the latest release. You'll need Claude Desktop installed first.

2. Double-click & connect

Double-click the file, click Install, and fill in the short form that appears:

Not sure what goes where? Ask whoever runs Laserfiche at your organization — it takes them a minute.

3. Ask

Open a chat and try:

• "Find every invoice from March in the Accounting folder."
• "What's in the Onboarding folder? Summarize the newest document."
• "Search for contracts mentioning Acme and list them with dates."

[!NOTE] Claude can look, but never change or delete — the extension is read-only by default, and your password lives in your operating system's keychain, not a text file.

Full walkthrough for end users and team rollouts: docs/desktop-extension.md.

For developers — the Python package

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

Requires Python 3.10+ and a reachable Laserfiche Repository API Server (self-hosted) with a service account that can read it, plus any MCP client (Claude Desktop, Claude Code, MCP Inspector). For local development:

git clone https://github.com/SamuelSHernandez/laserfiche-mcp
cd laserfiche-mcp
uv sync --extra dev

Configure

Copy the example file and fill in your repository details:

cp .env.example .env
$EDITOR .env

Minimum required variables for self-hosted password-grant auth:

Optional write-mode variables (fences and allowlists default off; the delete batch cap and required-field validation default on — see the Safety model section for context):

See .env.example for the full list including OAuth config, pagination limits, request timeout, retry attempts, and SSL verification.

API version note: LFRepositoryAPI ships with different routing surfaces across builds. Older self-hosted installs expose /v1/... paths; newer ones expose /v2/.... Probe your server with:

Terminal

Copy

curl {LF_REPO_API_URL}/v1/Repositories
curl {LF_REPO_API_URL}/v2/Repositories

Whichever returns a 200 with a JSON repo list is your version. If the wrong value is set, every call fails with 400 UnsupportedApiVersion. The default is v1 because that is what most current on-prem installations expose.

Auth note: Laserfiche self-hosted does not accept HTTP Basic auth. The server exchanges your username/password for a bearer token at POST /{api_version}/Repositories/{repository_id}/Token on first request and refreshes it automatically before expiry. The same flow works on both v1 and v2.

Connect to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "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"
      }
    }
  }
}

Restart Claude Desktop. The Laserfiche tools will appear in the tool picker.

Connect to Claude Code

claude mcp add laserfiche -- uvx laserfiche-mcp

(Pass env vars via --env LF_REPO_API_URL=... flags or set them in your shell before running Claude Code.)

Test it locally with the MCP Inspector

npx @modelcontextprotocol/inspector uvx laserfiche-mcp

This opens a UI where you can call each tool directly and watch the JSON-RPC traffic — useful for verifying endpoint shapes against your specific Repository API Server version before wiring it into Claude.

Remote HTTP (web clients)

The default transport is stdio — for local clients that launch the server as a subprocess (Claude Desktop, Claude Code, Cursor, Gemini CLI). Web and cloud clients (claude.ai custom connectors, ChatGPT connectors) can't spawn a local process; they connect to a URL. The same server can serve those clients over Streamable HTTP:

laserfiche-mcp --http                 # binds 127.0.0.1:8000, path /mcp
laserfiche-mcp --http --port 9000     # override port for this run

Configuration (all optional, LF_* env like everything else):

The --http server chooses its auth by precedence: OAuth (if LF_HTTP_OAUTH_ISSUER is set) → static token (if LF_HTTP_AUTH_TOKEN) → none (loopback only). OAuth is the multi-user path for claude.ai / ChatGPT; see Per-user OAuth below.

Verify locally with the Inspector (point it at the URL, not the command):

laserfiche-mcp --http &
npx @modelcontextprotocol/inspector   # then connect to http://127.0.0.1:8000/mcp

Per-user OAuth

For a multi-user connector, run the server as an OAuth 2.1 Resource Server — each user signs in through your existing identity provider (LFDS, Microsoft Entra, Okta, Auth0, Google) and the server verifies their token:

pip install 'laserfiche-mcp[oauth]'

LF_HTTP_OAUTH_ISSUER="https://login.microsoftonline.com/<tenant>/v2.0" \
LF_HTTP_PUBLIC_URL="https://lf.example.com/mcp" \
LF_HTTP_OAUTH_AUDIENCE="api://laserfiche-mcp" \
LF_HTTP_OAUTH_REQUIRED_SCOPES="laserfiche.read" \
  laserfiche-mcp --http --host 0.0.0.0

The server then serves protected-resource metadata (RFC 9728), so claude.ai / ChatGPT discover your authorization server, run the authorization_code + PKCE flow, and present a bearer token that this server verifies (signature via JWKS, plus aud / iss / exp / scopes). This is authentication at the edge — verified requests still reach Laserfiche via the shared service account, so the Laserfiche audit trail shows that account, not the end user. Full details, including IdP registration and the security checklist, are in docs/remote-http.md.

Connecting a web client

claude.ai and ChatGPT connectors need a public HTTPS URL. In practice that means putting this server behind a reverse proxy (or a tunnel like cloudflared / ngrok for a spike) and adding the resulting https://…/mcp URL as a custom connector in the client's settings.

[!WARNING] Read this before exposing --http to a network.

• Configure auth: OAuth (LF_HTTP_OAUTH_ISSUER) for multi-user, or at least a long random LF_HTTP_AUTH_TOKEN. Binding off-loopback with neither logs a warning and leaves your repository reachable unauthenticated.
• Terminate TLS at a reverse proxy in front of this server (it speaks plain HTTP).
• Your Laserfiche server is self-hosted behind a firewall; a public connector needs a deliberate network path in to it (VPN / DMZ / tunnel).
• See docs/remote-http.md for the full deployment and security checklist.

Command line (no model, no tokens)

Everything the MCP tools do against the repository, the same binary does from a shell — same code path, no LLM in the loop. Use it for the work that has one correct answer: inventories, downloads, dedupe, grepping a contract.

laserfiche-mcp setup                           # one-time: save connection details
laserfiche-mcp ls 1                            # list a folder (alias: list)
laserfiche-mcp ls '\HR\Leases'                 # ...or by path
laserfiche-mcp get 4821 --to ./lease.pdf       # stream a document to disk
laserfiche-mcp cat 4821                        # extracted text on stdout
laserfiche-mcp cat 4821 --pages 4-9            # just those pages
laserfiche-mcp find 4821 'unpaid balance'      # grep one document
laserfiche-mcp search 'unpaid balance'         # grep the whole repository
laserfiche-mcp manifest 1 --out inventory.csv  # walk a tree, write CSV
laserfiche-mcp dedupe 1                        # byte-identical documents
laserfiche-mcp diff 4821 4822                  # compare two entries (alias: compare)

Office-friendly synonyms work everywhere: list, read (cat), download (get), compare (diff), duplicates (dedupe). On Windows, quote paths with double quotes: laserfiche-mcp list "\HR\Leases".

ENTRY and FOLDER accept a numeric entry ID or a repository path. Every command takes --json, so the same invocation backs a shell pipeline:

laserfiche-mcp search 'termination' --json | jq -r '.results[].name'
laserfiche-mcp manifest 1 --json | jq '.by_extension'
laserfiche-mcp manifest 1 --out tree.jsonl --format jsonl && jq -r 'select(.extension=="pdf").name' tree.jsonl

Conventions that make these scriptable:

• Results go to stdout; progress and warnings go to stderr.
• Exit codes are meaningful: 0 success, 1 the operation failed or found nothing, 2 bad usage. laserfiche-mcp find 4821 'indemnify' >/dev/null && echo present does what you would expect.
• Nothing here writes to the repository. The subcommands are read-only by construction and never register the write tools.

What each command saves you

Document formats

cat, find, and the MCP's mode="text" read these with no extra dependencies: PDF, DOCX, PPTX, EML, HTML, RTF, and anything text/* (including CSV, JSON, XML, Markdown).

Two more need the optional extra:

pip install 'laserfiche-mcp[office]'   # adds .xlsx and Outlook .msg

Legacy binary .doc / .xls / .ppt are not supported — correct extraction needs an external converter such as LibreOffice, and bundling that would break the "pip install and it works" promise. Re-save as OOXML, or read Laserfiche's own indexed text with search.

Scanned images have no text layer at all. cat says so explicitly rather than returning an empty string, and points you at search, which reads the OCR index where that text actually lives.

Tools

Tool names below are shown in their original verb-first form (get_entry, set_fields, ...) for readability. In v2.0 every tool is also registered under the laserfiche_{resource}_{verb} form (laserfiche_entry_get, laserfiche_field_set, ...). Both names resolve to the same function. The laserfiche_* names are the recommended path; the old names remain as deprecation aliases through v2.x and will be removed in v3.0. The authoritative mapping lives in _V2_RENAME_MAP in src/laserfiche_mcp/server.py.

Reads (always registered)

Writes (registered only when LF_READ_ONLY=false)

Tools with two-step token return a preview + HMAC-signed confirmation_token on first call. Surface the preview to the user; on go-ahead, re-call with the same arguments plus the token. Tokens are bound to (operation, entry_id, entry_name) and the operation's own parameters (page_range for delete_pages, new_name for rename_entry, destination + name for move_entry) — executing with different arguments than were previewed fails verification. They expire after 5 minutes, and are invalidated by server restart (unless LF_CONFIRMATION_SECRET is set — see Configure).

Using search_natural

search_entries requires hand-written Laserfiche query syntax. If the server rejects the query the only feedback the LLM gets is a generic HTTP 400 — there's nothing actionable to retry against. search_natural is the LLM-friendly path:

1. First call — pass the user's question and (optionally) a folder_path to scope the answer; leave lf_query unset. The tool samples up to ten entries from that folder, returns the templates and field names it found, the Laserfiche search grammar reference, and 2–3 candidate query strings the LLM can choose from or refine.
2. Second call — same question, plus the chosen lf_query. On HTTP 400, the tool tries up to two automatic repairs (escape unescaped quotes inside values, then wildcard-wrap bare Name= values if fuzzy=True) before returning a structured error with all attempts visible so the LLM can author a fresh query.

The page-size cap for search_natural is the dedicated LF_MAX_PAGE_SIZE env var (default 100) — some self-hosted SimpleSearches implementations reject $top values above an internal limit, so this defaults lower than the list/folder cap.

Using search_content

The other search tools answer which entries matched. search_content answers what they say, by returning the matched passages themselves — page number, surrounding excerpt, and the exact substring that matched — pulled from Laserfiche's full-text index, which is where OCR output for scanned documents lives.

search_content(query="unpaid balance", folder_path="\Leases")
→ { "results": [
      { "entry_id": 7, "name": "lease-4821.pdf", "hit_count": 6,
        "hits": [ { "page": 4,
                    "text": "…tenant owes an unpaid balance of $2,400 as of March…",
                    "match": "unpaid balance" } ] } ] }

That answers most "what does this document say about X" questions without downloading anything. Reach for get_document_edoc(mode="text") only when an excerpt points you at the right document and you need more of it.

A bare phrase is wrapped into {LF:Basic~="..."} for you. Pass a query starting with { to use raw syntax instead — e.g. option="D" to search only OCR'd document text, or option="DFANLT" to permit leading and trailing wildcards.

Two knobs control cost: hits_for_top (how many results get their passages fetched — one extra request each, default 5) and hits_per_entry (passages per result, default 3, capped by LF_SEARCH_CONTEXT_HITS_MAX).

This is the asynchronous /Searches flow rather than SimpleSearches, which is what makes context hits available at all — they're keyed by a search token that only the async flow produces. Laserfiche caps concurrent searches per session (two, on v1), so the tool runs the whole create→poll→read→close cycle inside one call and always releases the token, including on timeout and failure. Older builds without the /Searches endpoints get a structured async_search_unavailable error pointing at search_entries as the fallback.

get_document_edoc modes

On v1 servers the Laserfiche Text export endpoint doesn't exist, so get_document_text cannot return anything. get_document_edoc gained a mode parameter as the workaround:

All tool descriptions are written to read like prompts — they tell the model when to use the tool, valid input shapes, and what kind of follow-up is expected. See src/laserfiche_mcp/server.py.

Troubleshooting

Start with one command:

laserfiche-mcp diagnose

It authenticates, probes every endpoint your Laserfiche build exposes, and prints an OK/unavailable table plus your write-mode and logging config. Failures are classified — it will tell you whether the server was unreachable (URL/VPN/TLS, not your password), whether the other LF_API_VERSION would work (it probes both and names the right one), or whether the credentials themselves were rejected.

No config yet, or config in doubt? Run the wizard:

laserfiche-mcp setup

It asks for the server URL, repository, and service account, saves them to ~/.laserfiche-mcp/.env (%USERPROFILE%\.laserfiche-mcp\.env on Windows), and ends with a diagnose run. Every later CLI invocation finds that file automatically when nothing else is configured.

Windows notes:

• In PowerShell/cmd, quote repository paths with double quotes: laserfiche-mcp list "\HR\Leases". Avoid a trailing backslash before the closing quote — it escapes the quote.
• Claude Desktop logs live at %APPDATA%\Claude\logs\ (macOS: ~/Library/Logs/Claude/). Look for mcp-server-laserfiche.log.

Common misconfigurations:

Errors

Every tool returns a stable dict on failure instead of raising — so the LLM gets actionable, structured data instead of Error executing tool ....

{
  "mode": "error",
  "operation": "laserfiche_entry_delete",
  "kind": "not_found",
  "error": "not_found",
  "status_code": 404,
  "server_error_code": null,
  "server_message": null,
  "reason": "Server returned 404 — the entry, path, or endpoint does not exist.",
  "request_id": "9f2c…",
  "upstream_trace_id": null,
  "entry_id": 999
}

kind is one of five canonical ToolErrorKind values — LLMs branch on this for category-level decisions (retry vs ask user vs abort):

error is the more-specific subkind. Server-mapped subkinds:

Tools also have pre-server mode: error shapes (path_not_allowed, path_traversal_blocked, exceeds_batch_cap, invalid_confirmation_token, missing_required_fields, page_range_required, invalid_page_range, invalid_name, invalid_field_name, invalid_tag_name, invalid_template_name, invalid_link_type, file_not_found, size_exceeds_cap, tool_not_allowed). list_repositories returns mode: fallback instead of erroring when the server doesn't expose the endpoint — see the docstring for the response shape.

See docs/error-contract.md for the full taxonomy, per-tool triggers, and the kind ↔ subkind mapping.

Safety model

Wondering what Claude actually sees and where your document content goes? See Data handling & privacy — it covers the data flow, what leaves the machine, and how to scope a service account so sensitive folders are never exposed. The rest of this section is about the write-mode guards.

Writes are off by default. When you enable them (LF_READ_ONLY=false), the following guards are available — all independent, all opt-in except as noted:

• Path-prefix fences (LF_WRITE_PATHS_ALLOW, LF_WRITE_PATHS_DENY) — every write checks the entry's fullPath (or the parent's for creates) against the configured prefixes. Case-insensitive, deny wins over allow, both \ and / accepted. move_entry fences on BOTH source and destination paths so a token from an allowed source can't be replayed to land in a denied folder. Strongest single fence — recommended for any non-trivial deployment.
• Local import source fence (LF_IMPORT_SOURCE_DIRS, default unset) — import_document reads file_path off the MCP process's own filesystem; with this set, the resolved (symlinks included) path must fall inside one of the configured directories or the tool refuses with source_path_not_allowed. This is the source-side counterpart to the path-prefix fences above, which only cover the repository destination.
• Tool-level allowlist (LF_WRITE_TOOLS_ALLOWED) — restrict which write tools register at all. Example: merge_fields,merge_tags,assign_template for a metadata-only deployment that can't create or delete anything.
• Folder-delete batch cap (LF_DELETE_FOLDER_MAX_DESCENDANTS, default 50) — delete_entry on a folder with more immediate children refuses unless force_large_delete=true is passed alongside the confirmation token. The preview surfaces exceeds_batch_cap: true so the LLM can explain the size before re-calling. If the child-count probe itself fails (transient error), the delete fails CLOSED — child_count_probe_failed: true on the preview, and execute refuses with child_count_probe_failed regardless of force_large_delete — rather than assuming an unknown count is safe.
• Audit-reason requirement (LF_REQUIRE_AUDIT_REASON, default false) — when true, delete_entry refuses without an audit_reason_id. Use get_audit_reasons to enumerate valid IDs.
• Required-field validation (LF_VALIDATE_REQUIRED_FIELDS, default true) — assign_template lists FieldDefinitions, finds isRequired: true fields, checks them against what's on the entry and what's in the caller's fields=, and returns a structured missing_required_fields error before the PUT — instead of the server's opaque Multistatus response. [9039].
• Two-step confirmation tokens (always on for destructive ops) — rename_entry, move_entry, delete_entry, delete_edoc, delete_pages return a preview + HMAC-signed token on first call; execute on second call. Tokens bind to (operation, entry_id, entry_name) plus the operation's execute-relevant parameters (page_range, new_name, destination), so the execute leg cannot silently swap in different arguments than the user confirmed; they expire after 5 minutes. By default the signing key is random per-process, so a restart invalidates pending tokens; set LF_CONFIRMATION_SECRET to derive a stable key instead (tokens survive restarts and verify across instances sharing the secret — for multi-instance deployments).

Recommended starting config for write mode

"env": {
  "LF_READ_ONLY": "false",
  "LF_WRITE_PATHS_ALLOW": "\\Sandbox\\mcp-test",        // scope to a sandbox first
  "LF_WRITE_TOOLS_ALLOWED": "create_folder,import_document,merge_fields,merge_tags,assign_template,delete_entry",
  "LF_DELETE_FOLDER_MAX_DESCENDANTS": "10",
  "LF_REQUIRE_AUDIT_REASON": "false"                    // turn on once you have a workflow
}

Pre-create the sandbox folder by hand in the Laserfiche web client; the fence needs an existing parent to read its fullPath. Once smoke-tested, broaden the tool list — path scope is still the strongest fence regardless of which tools are registered.

Roadmap

• Server-side audit logging — sidecar file with rotation, capturing every write tool call with the authenticated user, target entry, and outcome.
• Cloud — Laserfiche Cloud support (signin.laserfiche.com JWT-signed client_credentials flow plus the api.laserfiche.com v2-only endpoint surface).
• v3.0 — Remove the verb-first deprecation aliases (get_entry, set_fields, ...). Only the laserfiche_{resource}_{verb} names remain.
• Stateless MCP (spec 2026-07-28) — the protocol core is now stateless: the initialize handshake and session IDs are gone, and cross-call state must live in server-minted handles passed as ordinary tool arguments. This server is already shaped for that world — the HMAC-signed confirmation_token is exactly such a handle (set LF_CONFIRMATION_SECRET so every instance can verify every instance's tokens), and the async-search token never crosses a call boundary. The remaining review item for a remote/multi-instance deployment is that the per-process lifespan client and schema caches become per-instance. The cacheable tools/list in the new spec also raises the value of a small catalog (LF_LEGACY_TOOL_NAMES=false).
• Beyond — Workflow trigger tools, MCP resource links for edocs, and per-viewer table summaries for spreadsheet entries.

Development

uv sync --extra dev
uv run pytest                  # mocked HTTP, enforces 85% coverage baseline
uv run ruff check src tests
uv run mypy src

Tests use pytest-httpx to mock the Repository API and committed fixture PDFs to exercise the text-extraction paths — they don't require a real Laserfiche server.

Opt-in integration tests

LF_INTEGRATION_TEST=1 uv run pytest tests/test_integration.py

Reads the same LF_* env vars the server uses at runtime. Optional overrides:

• LF_INTEGRATION_FOLDER_PATH — folder used in the search_natural Mode A test (defaults to repository root)
• LF_INTEGRATION_PDF_ENTRY_ID — known PDF entry; if unset, edoc tests skip
• LF_INTEGRATION_SAFE_QUERY — a query expected to return results on your repo (defaults to {LF:Name="*"})

Use this before tagging a release if you have a reachable repository — it catches issues that mocked HTTP can't surface (server-side query syntax quirks, real PDF extraction, transport-level rejections).

Contributing

Issues and PRs welcome — particularly:

• Endpoint corrections for Repository API Server builds the v1 / v2 wire format hasn't been validated against
• Laserfiche Cloud client + JWT-signed client_credentials assertion flow
• Server-side audit logging for write-mode deployments (sidecar file + rotation)
• Text extraction for more document formats (ops/extract.py)

This is a community project, not affiliated with or endorsed by Laserfiche.

License

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