MCP

Wikai runs a Model Context Protocol (MCP) server so agents — the built-in chat sidebar, or an external tool like Claude Code in a terminal — can read and work with your wiki through a defined set of tools.

What it exposes

Reading: list and read pages, get metadata and labels, follow inbound and outbound links, search by tag or type, semantic concept search, regex grep, read images, list and read work documents, run a lint pass, and see what page you’re currently viewing.

Writing: create pages, update page bodies (respecting each page’s agent_edit policy), link text within a page, create work documents, and — when internet search is enabled — harvest web pages into the wiki.

Editing a single block: read, update, delete, or insert before/after one block rather than rewriting the page. The block keeps its id, so links and embeds pointing at it keep resolving — which a whole-page rewrite can’t promise.

Connecting an external agent

The MCP server runs locally, bound to 127.0.0.1 (your machine only), and starts with the app — there’s no toggle and no network exposure. Find the endpoint under Settings → Agents → MCP server, which shows the full URL — http://127.0.0.1:<port>/mcp — with a button to copy it. Point an MCP-aware client at that URL; for Claude Code, add Wikai to your mcpServers configuration with it.

The port changes each time Wikai starts. It’s assigned fresh on every launch, so if you’ve saved the URL in an external client’s config, update it after a restart — Settings → Agents always shows the current one. A future release will let you pin a fixed port, so an external client can be configured once and left alone.

Because it’s localhost-only, there’s no authentication in this version. If Wikai ever exposes the endpoint beyond your machine, token auth will come with it.

Last updated 2026-08-09.