What you need

• A Mac with macOS 13 or newer and the Terminal app (in Applications, Utilities).

• A Claude account on Pro, Max or Team. The free plan does not include Claude Code.

• Zimezumi in your Applications folder, with a vault open.

• Kendoshi: either the Mac app in Solo mode, or an account on a Kendoshi Team server.

• Two rules for everything below: never type paths and tokens by hand, always copy them. And tokens are passwords: they belong in no chat and no email.

Install Claude Code

Open Terminal, paste the line, press Enter. The installer puts Claude Code under your user account and keeps it up to date by itself.

curl -fsSL https://claude.ai/install.sh | bash

Close Terminal once and open it again, then check:

claude --version
claude doctor

The first line prints a version number, the second checks the installation and reports warnings with a suggestion. Then sign in: type claude, a browser opens, sign in with your Claude account there, come back to Terminal. /exit ends the session.

Connect Zimezumi

The MCP server is built into Zimezumi. There is nothing to download, no key and no Node. Zimezumi hands you the finished configuration with the real path to your vault.

1. Open Zimezumi and choose the vault Claude should work with.

2. Settings, section AI & MCP, click Check connection. Vault, installation, connection, tools and version must all be green.

3. Copy configuration. Your clipboard now holds this block; only the vault path is yours:

{
  "mcpServers": {
    "zimezumi": {
      "command": "/Applications/Zimezumi.app/Contents/Resources/mcp/zimezumi-mcp",
      "args": ["--vault=/Users/you/Documents/My Vault"]
    }
  }
}

Paste the command below into Terminal. The path to the server is the same for everyone. Replace only the value after --vault= with yours from the copied configuration.

claude mcp add zimezumi --scope user -- \
  /Applications/Zimezumi.app/Contents/Resources/mcp/zimezumi-mcp \
  "--vault=/Users/you/Documents/My Vault"

claude mcp list

claude mcp list now shows zimezumi. --scope user means: the connection applies in every folder where you start Claude Code, not only the current one.

Connect Kendoshi

Kendoshi has two modes, and the connection differs. Choose only the way that fits you.

Solo: the Mac app, data in your iCloud

There is no server for Claude to call. So the Mac app brings its own small server.

1. Open Kendoshi, Settings, expand the section API token (MCP & agents).

2. Switch on MCP server on this Mac. The token is created by itself and stored in the keychain.

3. Under Claude Code in the terminal click Copy. Your clipboard now holds the finished line.

4. Terminal, paste the line, Enter. Then claude mcp list, which now shows kendoshi as well.

Team: a Kendoshi Team server

1. Sign in to the team server in the browser, Settings, API token, Create token.

2. Copy link. The connector link looks like this: https://<your-server>/api/mcp?token=kdo_... The part after token= is your token. It is shown only once, so copy it straight away. Later you can only create a new one, which makes the old one invalid.

3. Paste the command below into Terminal, taking the server address and the token from the link.

claude mcp add --transport http kendoshi "https://<your-server>/api/mcp" \
  --header "Authorization: Bearer <token from the link>"

claude mcp list

The server also accepts the token directly in the link. For Claude Code the --header form is still the better one, because the token then does not end up in addresses and logs. For Claude Desktop or claude.ai in the browser use the link as it is: Settings, Connectors, Add custom connector, paste the link. No installation needed.

Check the connection: read first, then write

Start a new session: in Terminal, change to any folder and type claude. In the session type /mcp; both servers must show as connected. Then these three sentences one after the other, in this order:

Can you access Zimezumi? Tell me the vault and the life areas without changing anything.

What is on for me in Kendoshi today? Read only.

Create a note "MCP test" in the Zimezumi INBOX: one heading, one paragraph, one open checklist with two items.

The note appears in Zimezumi's sidebar by itself. If all three work, the technical part is done. What is still missing is the way of working.

Give Claude its working rules

Claude Code reads the file ~/.claude/CLAUDE.md at every start. What is in there applies in every session, in every folder. That is exactly where the rules belong: which app is for what, that Claude looks things up first, that tasks are created directly, and what a note in Zimezumi should look like. Then Claude no longer asks whether it should use the tools; it uses them.

This is how you create the file: in Terminal type mkdir -p ~/.claude && open -e ~/.claude/CLAUDE.md. TextEdit opens an empty file. Paste the block below, save, done. If the file already exists, append the block at the end.

# Working rules: Zimezumi and Kendoshi

## Roles
- Zimezumi is my knowledge: notes, documentation, decisions, results.
- Kendoshi is my work: tasks, dates, status.
- Both are connected via MCP. Use the tools directly. Do not ask me
  whether you should use them.

## Before you work
- For every topic, first search Zimezumi (search_notes) and read the
  matching notes (read_note). Never work from memory.
- For questions about tasks, first look in Kendoshi (my_work, day_context,
  get_board).

## Writing in Zimezumi
- Call get_editor_capabilities once per session. After that always use the
  blocks field, never html.
- Before update_note and append_to_note always read_note, and pass the mtime.
- Before creating a new note, search whether the topic already exists. If
  yes: append at the end as an addendum with the date in the heading. No
  duplicates.
- New notes go into the fitting folder, otherwise INBOX. Set tags and colour.
- Without asking: create a note, append an addendum, create a task.
- Ask first: replace a whole note, rename, move, change many notes at once.

## Formatting in Zimezumi, always, without being asked
- A status card at the top (data_card): date, status, what it is about.
- Then a key-message block (knowledge_block, kind info), two to four
  sentences, the most important thing first.
- Sections as headings (heading, level 2). Instructions as step headings
  (heading with step "1", "2", ...).
- Hints as knowledge_block with title: note for side remarks, info for
  context, success for things done, warning for traps and risks.
- Mark important sentences with highlight: yellow important, green decided,
  pink open. States as badge (success, warning, danger, info).
- Tasks and acceptance checks as task_list. Comparisons as table. Commands,
  prompts and code as code_block. References to other notes as
  internal_note_card.
- Short paragraphs, one thought per paragraph. No walls of text.
- No em dashes in the text. Comma, colon or full stop.
- At the end of every larger note: related notes and a changelog as a table
  (date, what).

## Working in Kendoshi
- Tasks that come up in conversation: create them at once (create_task), in
  the fitting project and group. Do not just mention them in the chat.
- More than three steps: create subtasks (create_subtasks).
- Keep the status current: in_progress at the start, feedback when I should
  check, done at the end.
- Deleting is not possible. Set finished things to done.

## Answers
- Key message first, then details. Short sections.
- Every answer ends with the next step or the open question.
- At the end of a task: list what was created or changed where
  (Zimezumi note with path, Kendoshi task, file).

The first real session

This prompt checks both connections, shows every building block once in a note, and creates the first task. Afterwards you know whether everything works, and you have a sample note to compare against. Start claude in Terminal, paste the block, Enter.

We are checking your connection to Zimezumi and Kendoshi. Follow the
working rules from CLAUDE.md.

1. Zimezumi: call get_vault_info and get_editor_capabilities. Tell me the
   vault path and the life areas. Change nothing.
2. Kendoshi: call list_projects and my_work. Tell me the projects and what
   is on today. Change nothing.
3. If both work: create a note "MCP test" in the Zimezumi INBOX that shows
   every building block once: status card, key message, two step headings,
   one hint each of note, info and warning, a checklist, a table, a code
   block, a yellow highlighted sentence and a changelog. Use the blocks
   field.
4. Create a task in my Kendoshi inbox: "MCP test: proofread the note in
   Zimezumi", due tomorrow.
5. Report back: what works, what is missing, and what you created (note
   path, task id).

If a tool is missing: tell me exactly which one and which command I should
run in Terminal.

Then open the note MCP test in Zimezumi. If it looks like this one, with a card at the top, coloured hints and numbered steps, Claude has understood the rules. If it looks like a bare block of text, CLAUDE.md is in the wrong place or the session was started before you saved.

Prompts for everyday use

Short and in plain language. Through the working rules Claude knows where to look and where to write.

What Claude cannot do, and why that is good

• In Zimezumi: delete nothing, unlock nothing, create no life areas. Before every change Zimezumi re-reads the note and stops if you changed it by hand in the meantime.

• In Kendoshi: delete nothing, not even into the trash. Claude has exactly your rights.

• Through the connection Claude also reads text that other people wrote: comments, wiki pages, forwarded mails. If such a text contains an instruction, Claude may take it for your request. Do not confirm proposals you did not ask for.

• With a cloud model, requested content travels to the provider. The local MCP server does not make the model local. What must not leave belongs in a locked note or in another vault.

When something is stuck

Changelog