Morphik MCP Server – Document Search for AI Assistants

TL;DR

TL;DR: Morphik publishes an official MCP server (@morphik/mcp) that lets Claude Desktop and other MCP-compatible AI assistants search a Morphik knowledge base, ingest files, and retrieve documents — without writing any retrieval code.

Source and Accuracy Notes

This section is MANDATORY. All links verified from actual source, not guessed.

• Project page: morphik.ai — verified June 28, 2026
• Source repository: github.com/morphik-org/morphik-core — README read June 28, 2026
• MCP documentation: docs.morphik.ai/using-morphik/mcp — verified June 28, 2026
• License: Business Source License 1.1 (verified via LICENSE file in repo)
• HN launch threads: github.com/morphik-org/morphik-core #43763814, #43601495

What Is Morphik?

Morphik is an AI-native document search and storage platform. According to its GitHub README, it provides “the most accurate document search and store for building AI apps” — with a focus on multimodal data that goes beyond plain text extraction. Morphik uses ColPali (a vision-language model approach) to understand the visual layout of documents, not just their text content.

The platform ships a hosted service with a free tier, plus a fully self-hostable open-source version.

Morphik’s core value proposition: traditional RAG pipelines break down on visually rich documents (charts, diagrams, tables). Morphik’s multimodal approach keeps those spatial relationships intact during retrieval.

What Is the Morphik MCP Server?

The @morphik/mcp package bridges Morphik to any AI assistant that supports the Model Context Protocol (MCP). Once configured, Claude Desktop (and other MCP clients) can:

• Search your Morphik knowledge base with natural language queries
• Ingest files from disk or base64-encoded bytes
• Retrieve full documents or individual relevant chunks
• Manage documents — list, delete, check ingestion status, set metadata filters

The MCP server speaks two transport protocols: StdIO (default, for local clients like Claude Desktop) and Streamable HTTP (for remote or web-based deployments).

Prerequisites

• Node.js 16.0.0 or higher
• A running Morphik server (local or cloud). Cloud users get a URI from morphik.ai/signup.

Setup Workflow

Step 1: Install the MCP Package

No explicit npm install step is needed for StdIO mode — npx handles it on first run. For Streamable HTTP mode:

npm install && npm run build:streamable

Step 2: Configure Claude Desktop

Open your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the morphik entry under mcpServers:

{
  "mcpServers": {
    "morphik": {
      "command": "npx",
      "args": [
        "-y",
        "@morphik/mcp",
        "--uri=your-morphik-server-uri"
      ]
    }
  }
}

For a local Morphik server, omit --uri or use --uri=local:

{
  "mcpServers": {
    "morphik": {
      "command": "npx",
      "args": [
        "-y",
        "@morphik/mcp"
      ]
    }
  }
}

Step 3: Restart Claude Desktop

Close and reopen Claude Desktop. The Morphik MCP tools become available in the conversation.

Available MCP Tools

Once connected, these tools are registered with the MCP client:

Ingestion

• ingest-text — Push raw text with optional metadata and folder scopes
• ingest-file-from-path — Upload a file from disk (sandboxed by --allowed-dir)
• ingest-file-from-base64 — Upload client-provided bytes (useful in Streamable mode)
• ingest-files-from-paths — Batch upload multiple files

Retrieval

• retrieve-chunks — Returns the most relevant text/image pages with optional padding, folder scopes, and metadata filters
• retrieve-docs — Returns whole documents instead of individual chunks
• get-pages-in-range — Fetches up to 10 consecutive pages from a document

Document Management

• list-documents — Paginated listing with folder/end-user scope support
• get-document — Retrieve metadata for a specific document ID
• check-ingestion-status — Poll processing state for a document
• delete-document — Remove a document and its derived data
• morphik-filters — View, set, or clear typed metadata filters (eq/regex/number_range/date_range)

File Navigation

• list-allowed-directories, list-directory, search-files, get-file-info — Help Claude browse only explicitly allowed directories when ingesting files.

Streamable HTTP Mode

If StdIO is not available (e.g., remote server, web deployment), use Streamable HTTP instead:

node build/index_http.js --uri=<your-uri> --allowed-dir=~/Documents

The server exposes:

• /mcp — JSON RPC requests
• /health — Health check endpoint

Deeper Analysis

Why this matters: The MCP protocol is rapidly becoming the standard way to give AI assistants structured, tool-mediated access to external data. Rather than building a one-off retrieval endpoint, publishing an MCP server means any MCP-compatible client can connect to Morphik — not just Claude.

Self-hosted vs. cloud: The open-source version gives you full control over the retrieval infrastructure. The cloud version handles scaling and management. Both use the same MCP interface.

Multimodal search: Morphik’s use of ColPali means it can search by visual understanding of documents, not just text extraction. This is particularly valuable for technical documentation with diagrams, screenshots, or mixed-media content.

Practical Evaluation Checklist

• Does Claude Desktop connect to the MCP server on restart?
• Can you ingest a PDF and retrieve relevant chunks with a natural language query?
• Do metadata filters correctly scope retrieval results?
• Does Streamable HTTP mode work behind a reverse proxy?
• Does the --allowed-dir sandbox correctly block access outside the specified directory?

Security Notes

• The --allowed-dir flag (StdIO) and allowed-dir parameter (HTTP) sandbox file access. Only directories explicitly passed are accessible.
• In Streamable HTTP mode, the server should be deployed behind TLS termination — the /mcp endpoint has no built-in authentication (MCP handles auth handshake).
• Claude Desktop config files should not be world-readable on shared systems.

FAQ

Q: Does this work with Cursor or other MCP-compatible editors? A: Yes. Any tool that implements the MCP StdIO or HTTP transport can connect, not just Claude Desktop.

Q: What file types does Morphik ingest? A: Morphik processes PDFs, images, videos, and other multimodal documents. The exact supported formats are documented at docs.morphik.ai.

Q: Can I self-host the Morphik server? A: Yes. The open-source version runs via Docker or direct installation. See docs.morphik.ai/getting-started. Note that self-hosted deployments have limited official support — the Discord community helps.

Q: What is the Business Source License? A: Morphik uses BSL 1.1, which permits free use for development and self-hosting, but converts to the Morphik Commercial License under certain conditions (e.g., running it as a service for third parties). See the LICENSE file for details.

Q: What Node.js version is required? A: Node.js 16.0.0 or higher.

Conclusion

The @morphik/mcp package is a clean, standards-compliant way to give AI assistants live access to a document knowledge base. With ColPali-powered multimodal search under the hood, it handles visually rich content that plain text extraction would destroy.

If you already use Claude Desktop or any MCP-compatible AI tool and have documents to search, this is a five-minute setup that replaces an entire RAG pipeline.

Get started at morphik.ai/signup or self-host via the open-source repo.