Groupdocs Search MCP

GroupDocs.Search MCP Server

MCP server that exposes GroupDocs.Search as AI-callable tools for Claude, Cursor, GitHub Copilot, and other MCP agents. It indexes folders of PDF, Office, email, archive and 70+ other document formats locally, then answers full-text, boolean, phrase, fuzzy and date-range queries over them — documents never leave your machine.

-6f42c1)

Installation

One-click installs pre-fill every supported setting: edit the placeholder host path in

the volume mount after install; an empty GROUPDOCS_LICENSE_PATH runs in evaluation mode.

More clients - ready-made configs for Claude Code, Codex CLI, Visual Studio 2022,

Cursor, Windsurf, Cline, and JetBrains Rider live in install/generated/.

This product ships via Docker (GHCR) only. The GroupDocs.Search engine DLL is about 200 MiB and ships ONNX Runtime natives, so the packed tool (~281 MiB) exceeds NuGet.org's 250 MB limit — dnx and dotnet tool install are not available for it. When a slimmer engine packs under the limit, the standard NuGet flow returns (see the revert notes in .github/workflows/publish_prod.yml).

docker run --rm -i \

-v $(pwd)/documents:/data \ ghcr.io/groupdocs-search/search-net-mcp:latest

Images are multi-arch (linux/amd64 + linux/arm64), tagged latest plus an immutable version tag per release (e.g. :26.10.1). To pin, replace :latest with the version tag — recommended for shared configs and CI.

Linux: run the container as your user

The image runs as a non-root user (uid 1654). On Linux a mounted folder keeps its host owner, so a folder you own is read-only inside the container and indexing fails with *"The index folder … is not writable"*. Add --user so the container writes as you:

docker run --rm -i --user "$(id -u):$(id -g)" \

-v $(pwd)/documents:/data \ ghcr.io/groupdocs-search/search-net-mcp:latest

MCP client configs (JSON) do not run a shell, so put the numbers in directly — id -u and id -g print them, typically 1000: "args": ["run", "--rm", "-i", "--user", "1000:1000", "-v", …]. Docker Desktop on Windows and macOS does not need this.

Where the index lives

Search is the one stateful server in the GroupDocs MCP family: an index is a persistent, searchable copy of your documents' text that survives restarts. By default it is kept in a .groupdocs-search-index folder inside the documents folder (/data/.groupdocs-search-index in the container), so the one volume mount above is all you need and the index is never indexed itself.

If the documents folder is read-only, keep the index in its own folder:

docker run --rm -i \

-v /path/to/documents:/data:ro \ -v /path/to/index:/index -e GROUPDOCS_MCP_INDEX_PATH=/index \ ghcr.io/groupdocs-search/search-net-mcp:latest

Available MCP Tools

Paths are always relative to the documents folder (contracts, reports/2026/q3.pdf), and indexes are referred to by name, never by filesystem path. Several named indexes can live side by side.

Query syntax (GroupDocs.Search DSL): single words (invoice); phrases in double quotes ("force majeure" — unquoted multi-word input does not match); AND, OR, NOT and parentheses; ? character wildcards (?ffect, princip?(2~4)); daterange(2026-01-01 ~~ 2026-06-30); number ranges (100 ~~ 500); field prefixes (filename: lease); regular expressions starting with ^; and fuzzy matching via the fuzzy argument.

Indexable formats include PDF, DOC/DOCX, XLS/XLSX, PPT/PPTX, ODT/ODS/ODP, RTF, TXT, HTML, EPUB, FB2, EML/EMLX/MSG, PST/OST, ONE and ZIP — including documents inside archives and email attachments. get_index_info with includeSupportedFormats: true lists the engine's full set.

Example prompts for AI agents

Copy any of these into Claude Desktop, Cursor, or GitHub Copilot Chat after the server is connected.

1. Make a folder searchable: *"Index everything in my contracts folder."*
2. Find the right documents: *"Which contracts mention a termination penalty?"*
3. Read the evidence, not the file list: *"Show me the paragraphs that talk about \"force majeure\", with the file each one comes from."*
4. Keep it current: *"I edited some files in reports — refresh the index and search for the Q3 forecast again."*
5. Separate collections: *"Create an index called hr for the HR folder, then search it for parental leave."*

Licensing

The MCP server itself is MIT; the underlying GroupDocs.Search engine requires a license for production use. Without one the server runs in evaluation mode, and the limits are counts, not watermarks — results look complete and are not:

• An index holds at most 100 documents (across all indexes of a repository); indexing stops there.
• A search returns at most 10 documents.
• Every tool response carries an evaluation note while this applies.
• Full list: evaluation limitations.

There are two ways to license the server.

License file — mount your GroupDocs.Search.lic (or GroupDocs.Total.lic) into the container and point GROUPDOCS_LICENSE_PATH at it. Works offline.

Metered (pay-per-use) — set both GROUPDOCS_METERED_PUBLIC_KEY and GROUPDOCS_METERED_PRIVATE_KEY. You are billed for what you actually process, which suits AI agents well because their usage is bursty and hard to predict up front. If both a metered key pair and a license file are configured, metered wins and the file is ignored.

Metered mode needs outbound connectivity. Usage is reported to GroupDocs servers, so air-gapped

or firewalled deployments must allow that egress — or use a license file instead.

Call the `get_license_status` tool at any time to see which mode is active, and — under metered — the consumed quantity and remaining credit. It processes no document and changes nothing.

Keeping the private key out of your config

The metered private key is a secret. MCP client configs are plain files on disk, and a project-scoped .mcp.json is committed by convention, so prefer, in order:

1. Set the variables in your OS environment and leave them out of the config entirely. An stdio MCP server inherits its client's environment, so this works in every client. On macOS note that an app launched from Finder does not inherit shell exports — use launchctl setenv, or launch the client from a terminal.
2. Reference them indirectly if your client supports it (Claude Code expands ${VAR}; VS Code uses ${env:VAR} and ${input:...}).
3. A literal value in a user-scoped config (for example ~/.claude.json) is acceptable for a single developer.
4. Never commit a literal key in a project-scoped .mcp.json.

For Docker, forward the values without writing them into any file — the names alone, so the keys never appear in the command line:

docker run --rm -i \

-e GROUPDOCS_METERED_PUBLIC_KEY -e GROUPDOCS_METERED_PRIVATE_KEY \ -v /path/to/documents:/data \ ghcr.io/groupdocs-search/search-net-mcp:latest

• Get a free 30-day temporary license
• Purchase a license | Product page

Configuration

| GROUPDOCS_METERED_PRIVATE_KEY | Metered priv