Setting Rusuban up from Claude

Everything you can set with rusuban commands, you can ask Claude Code to set. rusuban init registers Rusuban's MCP server for your user; any Claude Code session on this machine then has its tools.

What Claude can do

  • Read the instance: projects, their items and cards, runs, tasks, the journal.
  • Change a setting: add a repo, create or remove a project, a project's approval bypass, the host profile, the merge switch, the limits, the runs' model, the work root, the owner, a task, a channel, remote access.

Ask in your own words: "watch acme/widgets in Rusuban", "pause the merges", "write a task that drafts the release notes every morning".

What waits for you

A change that widens what agents may do, or where data goes, never applies from Claude. It waits as a card in the dashboard's inbox, under Rules & settings, marked Setting, with two answers: Allow it or Don't allow it. Claude tells you it is pending.

What widens:

  • a project added, or a repo added to one: agents act on a repo as its rules say;
  • a project's approval bypass turned on, the merge switch armed;
  • the host profile set to bypass;
  • a test machine added, since agents run code on it;
  • the limits raised, a new owner;
  • a task written or changed: its card shows the task, or the diff of its file;
  • a channel added, since cards leave for its target;
  • remote access turned on.

What narrows applies at once: the bypass off, the switch paused, the host profile set to allowlist, the limits lowered, a project, a channel or a task removed, remote access off. So does a project without a repo created: it starts nothing by itself, and each of its tasks waits for you. So does the runs' model: every run uses the one you chose, never Claude Code's default, and without one no run starts. So does the work root: the next workers get their work folders there.

In a terminal, rusuban asks you to type the change's target instead of making a card. Claude's own shell has no terminal: there, the change waits for the card too.

What Claude never does

  • Answer a card. Cards are for you: an agent would otherwise approve its own proposals.
  • Reach the settings from a Rusuban run. The agents Rusuban starts have the MCP server and the owner's commands denied, and both refuse to run inside a run.

Under the bypass host profile, an agent's shell can still get around these limits. The security model says how far they go.

Every change is journaled

Each setting goes to the journal with where it came from: a terminal, a command without a terminal, Claude through the MCP, or your answer in the dashboard. Ask Claude for the journal, or open it in the dashboard.