Nuxeo MCP: AI meets hyperscale content

There is a quietly beautiful thing about infrastructure that simply works. Code that opens doors rather than closes them. As one of the contributors to the Nuxeo MCP Server, I've watched this project grow from a quick scaffold into a working bridge between Nuxeo and AI. Today, I'm excited to share what it does, how it's built, and how *you* can help shape it into something even more powerful.

What Nuxeo MCP exposes

MCP means Model Context Protocol. It defines an interface between AI applications and external capabilities; it is not specific to Nuxeo. Maretha's server connects that interface to a Nuxeo repository through tools, resources and prompt templates. The repository documents document query, retrieval and modification tools. Nuxeo MCP project

Match the transport to the client

The server implementation inspected on September 10, 2026 uses standard input/output by default, Streamable HTTP with --http, and an SSE transport with --sse. Pin the repository revision and dependencies when reproducing a setup. These are source-checked options, not a claim that a deployment has been tested against your client. Server entry point

Streamable HTTP can itself use SSE for streamed messages. The distinction is therefore not simply "HTTP for short jobs, SSE for long jobs." The MCP transport specification defines the transport behavior; a long-running Nuxeo job still needs application-level submission, status and error handling.

Keep resources, prompts and tools distinct

Resources expose content, prompts provide reusable instructions, and tools perform callable operations. The project's prompt module currently registers a document-type search prompt through a decorator. Follow the real module when adding a prompt; the earlier register_prompt sketch was conceptual, not an executable project API. Prompt module

A summarization workflow should specify which document fields are retrieved, which identity may read them, and where the generated result may go. Test the retrieval step separately from the model's response.

Start with a bounded acceptance test

1. Choose a reviewed repository revision and a non-production content sample.
2. Configure the intended repository URL and a limited identity explicitly.
3. Connect a compatible client and inspect its available tools and resources.
4. Retrieve allowed content and verify that restricted content remains inaccessible.
5. Exercise a write operation only in a disposable test location, with approval controls appropriate to the client.
6. Record errors, latency and dependency behavior before expanding the scope.

Production concerns to resolve

Define the exposed operations, identity mapping, audit requirements and request limits. Authentication between the MCP server and Nuxeo does not by itself secure every network-facing MCP endpoint. Review that boundary explicitly.

Repository content can also contain misleading instructions. Treat retrieved material as data and keep authorization decisions outside model-generated text. Measure the value of the workflow alongside its operating cost.

Read more about Nuxeo AI integration, or discuss a bounded implementation with Maretha's integration team.