WikiMCP

What this repo gives you

• A Wiki.js stack (Wiki.js + Postgres) you can run locally with Docker.

• An MCP server (Node/TypeScript) that exposes Wiki tools over stdio (spawned by a client like LM Studio / Cursor / Claude Desktop).

• Receipts-first tooling: wiki.quote returns exact line ranges so answers can cite sources.

Architecture in one picture

Important: Docker is for Wiki.js + Postgres. The MCP server is typically spawned directly by your MCP client (no Docker required for the MCP server unless you want it).

Quick Start (sample data)

• Wiki.js opens at: http://localhost:3030

• The stack includes sample content in ./wiki-content

Setup paths (choose your adventure)

1) Local with sample data (default)

Uses ./wiki-content (characters, locations, chapters, etc.).

2) Local with empty wiki

Creates ./wiki-content-empty with just the en/ locale folder (sample data stays untouched).

3) Local with custom content path

Points to any folder on your machine.

4) Reset everything (wipe DB, keep content)

Removes Docker volumes and rebuilds from scratch. Your content folder survives.

Wiki.js first-run setup

After containers start:

1. Open http://localhost:3030

2. Complete the Wiki.js setup wizard (create admin account)

3. Go to Administration → Storage → Local File System (or Git)

4. Set the path to: file:///wiki-storage

5. Click Apply and Import Everything

Your markdown files will appear in the wiki.

Connect an LLM / MCP client (the fun part)

WikiMCP communicates via stdio (stdin/stdout). Your MCP client spawns the server directly.

1) Build the MCP server

This produces dist/index.js.

2) Configure your MCP client

Cursor / Claude Desktop (example)

Add to your MCP config (example file: mcp-config.example.json):

Adjust paths for your system.

LM Studio (concept)

LM Studio can act as an MCP host as well. Add a server entry that spawns:

• node dist/index.js …and pass WIKI_REPO_PATH to point at your content directory.

MCP vs REST/Swagger (1-minute version)

If you’re used to normal APIs, MCP can feel weird at first. Here’s the simple mapping:

• REST/Swagger: You design lots of URLs like GET /pages/{id} and document them in OpenAPI.

• MCP: You expose a tool list (functions) like wiki.search, wiki.get_page, wiki.quote with JSON schemas.

In other words:

• REST is resource-oriented (endpoints)

• MCP is capability-oriented (tools)

A typical REST call:

• GET /api/wiki/pages/Tim

A typical MCP call (conceptually):

• call tool wiki.get_page({ path: "/characters/white-bunny" })

The MCP client/host (LM Studio, Cursor, etc.) is the thing that:

1. discovers the tool list

2. decides when to call a tool

3. feeds tool results back into the model

MCP Authentication (optional) (optional)

If MCP_AUTH_USERNAME and MCP_AUTH_PASSWORD are both set, every tool call must include:

Bootstrap helpers:

or:

Write operations

wiki.create_page is disabled by default.

Enable it by either:

1. .env:

• set WIKI_WRITE_ENABLED=true

• ensure your wiki content mount is not read-only

or

2. bootstrap flag:

Available MCP tools

wiki.get_context_bundle (the power tool)

This is the bridge between librarian and editor. Instead of orchestrating multiple calls, ask:

Returns:

• mainPage: Full content of the best-match page

• relatedPages: Summaries of linked and related pages

• citations: Version info for every page (receipts)

• totalPagesFound: How many matches exist

The model can now say "give me everything about X" and get a ready-to-use context bundle.

Environment variables

Project structure

Troubleshooting

“password authentication failed”

Old Postgres volume credentials. Reset:

“detected dubious ownership in repository”

Git safety check. The compose file configures safe.directory. Restart:

MCP server not connecting

• Ensure npm run build succeeded

• Ensure absolute paths in the MCP client config

• Ensure WIKI_REPO_PATH points to real content

How this differs from a normal REST API (quick intuition)

WikiMCP is not a traditional HTTP API meant for humans or frontends.

• A REST API exposes endpoints (URLs) and HTTP verbs.

• An MCP server exposes tools (named capabilities with JSON schemas).

In practice:

• Your MCP client (LM Studio, Cursor, Claude Desktop) spawns this server.

• The server advertises its tools (e.g. wiki.search, wiki.quote).

• The LLM decides which tool to call and with what arguments.

• WikiMCP enforces validation, permissions, and returns structured results.

Think of it as "controllers for an AI client" rather than endpoints for a web app.

If you understand that one sentence, you understand MCP well enough to build useful things.

When not to use WikiMCP

WikiMCP is intentionally simple. You probably should not use it if:

• You only want a normal REST API over your wiki

• You don’t need citations or exact quotes

• You don’t plan to use an LLM with tool-calling

WikiMCP shines when:

• You want an LLM to prove where it got information

• You’re experimenting with agents, RAG, or MCP

• You want a clean learning project that scales into something real

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).

WikiMCP intentionally mirrors the license of Wiki.js, which is also AGPL-3.0. This ensures:

• Clear legal alignment with the upstream project

• Transparency for anyone running modified versions as a service

• A shared expectation that improvements remain open

If you are unfamiliar with AGPL, the practical takeaway is simple:

If you modify this project and make it available to users over a network, you must provide the source code of your modified version.

For personal use, learning, tinkering, and local deployments, AGPL imposes no additional burden.

See the LICENSE file for full terms.