wikicore

An MCP server for a wiki that your AI assistants keep for you: a git repo of Markdown notes (Karpathy's LLM-wiki pattern) that Claude, ChatGPT and coding agents all read and write, so what one of them learns the others know.

• Your assistant does the reasoning; wikicore does lookup and filing. No LLM calls, no embeddings, no database on the server.

• The vault is the instance. Page types (_schema/*.yaml), the rules every client receives (_instructions.md), tool descriptions (_tools.yaml) and optional modules (repos.yaml) all live in the vault. You customise by editing files, never by forking.

• Plain files. Markdown with YAML frontmatter, one folder per type, [[wiki-links]], generated index.md and append-only log.md. Opens as an Obsidian vault. Every change is a git commit authored by the assistant that made it.

Tools

Read: search, get, related, list_notes, recent, schema, status. Write: save_source (keep verbatim input: meeting notes, transcripts, paper text), file (create a page; refuses duplicates by title or alias), update (append to or replace a section, merge frontmatter), add_type. Setup: bootstrap (per assistant: file what it already knows from its memory and chat history). Repos (only when repos.yaml exists): repos, register_repo, repo_status, repo_read, repo_log, repo_tree.

Web: / connect guide, /login (OAuth, paste your key), /upload (files and PDFs into unfiled sources, up to 25 MB), /health.

Related MCP server: 50 First Tapes MCP Server

Run locally

uv sync
uv run wikicore init ./vault && git -C vault init -q
WIKICORE_VAULT=./vault WIKICORE_DEV_CLIENT=me uv run wikicore serve   # no auth, local only
claude mcp add --transport http wiki http://127.0.0.1:8050/mcp
uv run pytest

With auth: uv run wikicore token add owner prints a wk_ key. CLI agents send it as Authorization: Bearer wk_...; claude.ai and ChatGPT connect with OAuth and you paste the key once.

Commands

• wikicore init <dir>: create a vault from the defaults. On an existing vault it only adds missing default files and never overwrites your edits.

• wikicore serve: run the server (a single process; OAuth state is per process).

• wikicore check: report unreadable pages, schema warnings and dangling links. Exit code 1 if any.

• wikicore token add|list|revoke <client>: manage client keys. Revoking a key also ends the OAuth sessions it created.

• wikicore oauth revoke-all: end every OAuth session (client keys stay valid).

Customise

_tools.yaml:

search:
  description_append: "Use for anything about my research, the people in it and its funders."
  parameters:
    query: "Names, project names or topic words."
bootstrap:
  disabled: true

_instructions.md frontmatter name sets the name clients show; its body is sent as the MCP server instructions. Add a type with the add_type tool or by dropping a file in _schema/.

Server name and instructions are read at startup, so changes to _instructions.md need a restart (a deploy does that). Tool descriptions refresh when the vault syncs.

Environment

wikicore serve reads these variables.

Code repos

Put a repos.yaml in the vault to let assistants read your code repos:

repos:
  - name: analytical-engine
    url: https://github.com/example/analytical-engine
    project: analytical-engine   # optional: slug of a project page
    branch: main                 # optional

The repo tools are registered only when repos.yaml exists, and repos: [] enables them with no repos. Assistants can then add entries with register_repo. It refuses to run while any entry in the file is unreadable, so fix those by hand first. Adding the file needs a restart.

Deploy

Copy .deploy.env.example to .deploy.env, fill it in, run ./deploy.sh. It sets up a system user, /opt/<app>, a systemd service, an nginx vhost (uploads up to 26 MB) and a certificate, and prints the deploy key to add (with write access) to your private vault repo. Run one server process only.

License

MIT