Jahia MCP Community Server

GitHub: https://github.com/Jahia/jahia-mcp-community-server

An OSGi bundle that exposes Jahia's GraphQL API as a Model Context Protocol (MCP) server, enabling AI assistants such as Claude to query and mutate Jahia content directly over a secure HTTP endpoint.

How it works

The servlet registers at /modules/community-mcp and implements the stateless MCP JSON-RPC 2.0 protocol over HTTP. When an MCP client calls a tool, the servlet dispatches the GraphQL request in-process via an OSGi service reference to OsgiGraphQLHttpServlet — no extra HTTP hop is involved.

MCP Client (Claude Code)
    │  POST /modules/community-mcp
    │  Authorization: APIToken <token>
    ▼
McpServlet  ──[whitelist check]──►  OsgiGraphQLHttpServlet
                                     (graphql-dxm-provider)

The caller's Authorization header and the current Jahia user are forwarded so that all JCR permission checks apply normally.

Related MCP server: mcp-graphql-tools

Requirements

• Jahia 8.2+ with graphql-dxm-provider deployed

• personal-api-tokens module deployed (for API token authentication)

• Java 17

Installation

Deploy the bundle JAR via the Jahia Module Manager or drop it into $JAHIA_HOME/modules/.

On first activation the module automatically seeds a set of default skills into JCR under /sites/systemsite/contents/mcp-skills/.

Authentication

Access to /modules/community-mcp requires a personal API token with both the graphql and community-mcp scopes.

Generate one in Jahia under Administration → Profile → Personal API Tokens, or via GraphQL:

mutation {
    admin {
        personalApiTokens {
            createToken(name: "my-mcp-token", scopes: ["graphql", "community-mcp"])
        }
    }
}

Pass the token in every request:

Authorization: APIToken <your-token>

The community-mcp permission is automatically granted to users with the admin role (configured in META-INF/configurations/org.jahia.bundles.api.authorization-community-mcp.yml).

MCP client setup (Claude Code)

Add the following to ~/.claude/settings.json:

{
  "mcpServers": {
    "jahia-mcp": {
      "type": "http",
      "url": "http://localhost:8080/modules/community-mcp",
      "headers": {
        "authorization": "APIToken <your-token>"
      }
    }
  }
}

Available tools

introspectSchema

Returns all available top-level GraphQL query and mutation operations with full type details. Call this first to discover what operations and arguments are available before calling executeGraphQL.

No input arguments required.

executeGraphQL

Executes any GraphQL query or mutation against Jahia's graphql-dxm-provider, subject to the configured whitelist.

Example — list root child nodes:

query {
    jcr {
        nodeByPath(path: "/") {
            children {
                nodes {
                    name
                    path
                    primaryNodeType { name }
                }
            }
        }
    }
}

Introspection fields (__schema, __type) always pass regardless of whitelist configuration.

listSkills

Returns the name, display name (mcpName), and description of every skill stored in JCR under /sites/systemsite/contents/mcp-skills/. Skills can be organized in sub-folders at any depth.

No input arguments required.

Example response:

[
  {"name":"hello-jahia","mcpName":"Hello Jahia","description":"How to respond to Hello Jahia"}
]

getSkill

Returns the full Markdown content of a skill by name. Call listSkills first to discover available names.

Skills

Skills are Markdown documents stored as mcp:skill JCR nodes. They allow Jahia-specific knowledge and instructions to be delivered dynamically to any MCP client session.

Default skills

A set of default skills is seeded into JCR on module activation. They are defined as .md files under src/main/resources/skills/ and are only created if the node does not already exist.

Managing skills via GraphQL

All MCP operations are grouped under a single mcp namespace on Query and Mutation.

# List all skills
query {
    mcp {
        skills {
            name
            mcpName
            description
            content
        }
    }
}

# Create or update a skill
mutation {
    mcp {
        saveSkill(
            name: "my-skill",
            mcpName: "My Skill Display Name",
            description: "What this skill does",
            content: "# My Skill\n\nInstructions for Claude..."
        )
    }
}

# Delete a skill
mutation {
    mcp {
        deleteSkill(name: "my-skill")
    }
}

Adding a bundled default skill

Drop a .md file with frontmatter into src/main/resources/skills/<subfolder>/:

---
mcpName: My Skill Display Name
description: Short description of what this skill does
---
Full Markdown instructions here...

The file name (without .md) becomes the JCR node name. The folder hierarchy mirrors the JCR sub-folder structure under mcp-skills/.

Security notes

Allow-all mode (empty whitelist)

When the whitelist is empty (the default), all GraphQL operations are permitted to any caller who holds a valid API token with the community-mcp scope. The server logs a WARN on every request in this state to make it visible in logs:

WARN McpServlet - MCP GraphQL gate is running in ALLOW-ALL mode (whitelist is empty). ...

Configure an explicit whitelist in Administration → MCP Server (or in META-INF/configurations/org.jahia.community.mcp.cfg) to restrict which operations MCP clients may call.

Hardening recommendation: treat the empty (allow-all) whitelist as an out-of-box convenience, not a production posture. Because the community-mcp scope is granted to the admin role, a token in allow-all mode can run any GraphQL operation the admin user is permitted to (bounded only by JCR ACLs) — including administrative mutations. For least-privilege, always configure an explicit whitelist scoped to exactly the operations your MCP clients need.

Named fragment spreads blocked when whitelist is active

When a whitelist is configured, GraphQL queries that use named fragment spreads (...FragmentName) are rejected. The whitelist gate cannot resolve spread contents without the full fragment definition, so it fails closed rather than allowing a bypass.

Inline fragments (... on TypeName { ... }) are fully supported.

X-Forwarded-For not trusted

Client IP logged in audit/block messages is taken from HttpServletRequest.getRemoteAddr() (the TCP peer address). X-Forwarded-For is intentionally ignored to prevent IP spoofing. If Jahia is behind a trusted reverse proxy, configure RemoteIpValve at the Tomcat layer to rewrite remoteAddr to the real client IP before the request reaches the servlet.

Request body size cap

POST bodies larger than 2 MB are rejected with a JSON-RPC -32700 error before any parsing.

Access control — Allow List

The admin UI at Administration → MCP Server lets you restrict which GraphQL operations the MCP server may execute.

Behaviour

Dot-path entries

Entries are dot-separated field paths that mirror the GraphQL selection hierarchy:

Whitelist entries are persisted in the OSGi configuration org.jahia.community.mcp and can also be set manually in META-INF/configurations/org.jahia.community.mcp.cfg:

whitelist=jcr,currentUser,admin.jahia.isAlive

Blocked operation log

Every blocked operation is logged at WARN level with the operation path, the authenticated user, and the client IP:

WARN McpServlet - MCP operation blocked: path='admin.jahia.shutdown', reason=not in the whitelist, user='john', ip='10.0.0.5'

Health check

A GET /modules/community-mcp request returns a JSON status response for authenticated users with the community-mcp permission:

{"status":"Jahia MCP server running","version":"1.0.0","tools":["executeGraphQL","introspectSchema","listSkills","getSkill"]}

Testing

Docker-based Cypress integration tests live in tests/. They require a running Docker environment.

cd tests
cp .env.example .env          # fill JAHIA_IMAGE and JAHIA_LICENSE
yarn install
./ci.build.sh                  # build the Cypress Docker image
./ci.startup.sh                # start Jahia + run tests + collect results

Test results are written to tests/results/.

Test coverage