Sharing a weekend project: Connecting AI Agents to Wiki.js via MCP

[uyu423]

Hi everyone!

I’ve been a happy user and a sponsor of the Wiki.js project since 2023. Currently, I'm running my personal knowledge base at wiki.yowu.dev.

As AI agents (like Claude) have become a big part of my daily coding workflow, I recently started thinking about how to better preserve the "fragmented knowledge" that comes out of these AI conversations. Since Wiki.js has such a clean GraphQL API, I realized it would be the perfect backend for a Model Context Protocol (MCP) server.

So, I spent a few hours today whipping up a lightweight MCP bridge for Wiki.js. I wanted to share it with the community in case anyone else is looking for a similar workflow.

• Repo: uyu423/requarks-wiki-mcp
• NPM: @yowu-dev/requarks-wiki-mcp

Real-world Example: Documenting as I build

To give you an idea of how this works, I actually used this MCP server to document its own development process.

For instance, while I was working on the npm publish workflow for this project, I encountered a few specific issues. Instead of letting that knowledge disappear in my terminal history, I used the MCP server along with a custom "Knowledge Manager" skill I wrote for my agent. The result was this structured wiki page, generated almost instantly:

👉 https://wiki.yowu.dev/en/dev/npm/github-ci-publish

• The MCP server includes a wiki.js-specific Markdown syntax prompt.
• For reference, I wrote a structured SKILL like https://gist.github.com/uyu423/e963c6a9b844440b31fa8a99ee6852fd and use it with MCP.

Philosophy: Simple Bridge, Powerful "Skills"

I designed this MCP server to be a simple, transparent bridge that handles the GraphQL mutations and queries.

The real power comes from how you prompt your agent. Rather than hard-coding logic into the server, I found it much better to create separate "skills" (system prompts) that tell the agent how to categorize, path-name, and format the content for Wiki.js. From my testing, this approach is incredibly flexible and makes the documentation process feel frictionless.

Just a simple bridge

If you're into AI agents and want to try it out, here’s a quick setup for Claude Desktop:

{
  "mcpServers": {
    "requarks-wiki": {
      "command": "npx",
      "args": ["-y", "@yowu-dev/requarks-wiki-mcp@latest"],
      "env": {
        "WIKI_BASE_URL": "https://wiki.your-domain.dev",
        "WIKI_API_TOKEN": "your_wikijs_api_key_jwt",
        "WIKI_GRAPHQL_PATH": "/graphql",
        "WIKI_DEFAULT_LOCALE": "en",
        "WIKI_DEFAULT_EDITOR": "markdown",
        "WIKI_MUTATIONS_ENABLED": "true",
        "WIKI_MUTATION_CONFIRM_TOKEN": "CONFIRM_UPDATE",
        "WIKI_MUTATION_DRY_RUN": "false",
        "WIKI_ALLOWED_MUTATION_PATH_PREFIXES": "",
        "WIKI_HTTP_TIMEOUT_MS": "15000",
        "WIKI_HTTP_MAX_RETRIES": "2"
      }
    }
  }
}

(For detailed full mcp configuration, please read the https://github.com/uyu423/requarks-wiki-mcp README)

It supports searching, reading page content, browsing the page tree, and creating/updating pages with some built-in safety guards.

I'm just happy to contribute something back to the community of a project I’ve supported for years. If you use Wiki.js and AI tools, I'd love to hear your thoughts!

Huge thanks to @NGPixel and all the contributors for making Wiki.js awesome.

Cheers.

[chgjin]

very good

[Erudition]

Beautiful skill! I've incorporated many of the wikijs specifics into my skill (modified karpathy-wiki idea) for Hermes, since I'm using a git wiki to share as the agent's knowledge and a human-browsable directory.