Getting started

From nothing to a running team. If you have never used Multica, start at step 1 — the whole thing is a few minutes, and after the last step you never type a command again unless you want to.

What you need

• An agent CLI you already talk to — Claude Code, Cursor, Codex, Windsurf, Gemini CLI. Mops runs inside it.
• The multica CLI — the small binary that connects your machine to a Multica workspace. You install this one (cloud, self-hosted and the desktop app are different choices, so Mops points you at multica.ai rather than guessing); everything after it is Mops's job.

That's it. Once the CLI answers, Mops takes over — no account-setup marathon, no dashboard to configure.

1 · Install the skill

Pick the line for your harness:

# any agent, via skills.sh
npx skills add jamillazarev/multica-ops

# Claude Code as a plugin (adds slash commands)
claude plugin marketplace add jamillazarev/multica-ops
claude plugin install multica-ops@multica-ops

# into a Multica workspace, as an agent skill
multica skill import --url github.com/jamillazarev/multica-ops/tree/v0.4.15/skills/mops

All three routes were run end to end — measured, not assumed (the first two on 2026-07-31, the import re-measured 2026-08-01, each against the version current that day). What is measured is the route, never the tag in the line above — that moves with every release, and claiming a fresh measurement for a tag that did not exist when the run happened is exactly the kind of true-looking sentence this file is careful about. Four things worth knowing before you paste:

• The plugin route is two commands, two steps. In-session, /plugin opens the same thing as a menu: paste only the source — jamillazarev/multica-ops — into its Add Marketplace field, then install from the list. Both lines in that one field is an error it hands straight back.
• The import URL points at the folder that holds SKILL.md, not at the repository. Here that is skills/mops. The repository root answers "The Multica service is temporarily unavailable (server error)" — the CLI mislabelling a 502, which --debug shows for what it is. Measured both ways on 2026-08-01.
• Pick one route per harness, not both. The plugin manifest (.claude-plugin/) sits at the repository root while the corpus sits in skills/mops/, and a skills.sh install names the repository rather than a folder — so it copies the manifest too, Claude Code sees two plugins of one name, and the installed plugin wins while the copy reads "not loaded". No duplicate commands, and the copy still serves every other harness from ~/.agents/skills/; delete ~/.claude/skills/multica-ops if you want the message gone.
• ~/.agents/skills/multica-ops is load-bearing even if you never chose it. Found on a real machine 2026-08-01: Factory and Pi both install as a symlink into that path, and if no route ever created it, both links are broken while looking installed — the directory listing shows the name, and nothing resolves. scripts/find-installs.sh reports them as BROKEN, which is how these two were found. If you use either harness, put the repository there (a clone, or rsync -a --delete --exclude .git) and the links come alive.

In Claude Code the commands arrive namespaced — /multica-ops:init, /multica-ops:status, and /multica-ops:mops <anything> as the free-text front door. The prefix is how plugin commands always work: it is what stops two plugins colliding over one verb. Everywhere else there are no slash commands and none are needed — plain language reaches every flow.

2 · Just say what you're making

You don't need to know a command. Say what you want — "set up a team for my Android app", "I have a Linear backlog, move it here" — or type /multica-ops:mops on its own, or even "hi".

Mops runs "day zero" for you — six checks, reported as one list with the fix for anything missing, not six separate prompts:

Check	If it's missing
Multica installed?	This one is yours. Mops points you at multica.ai — cloud, self-hosted, or the desktop app (it bundles the CLI) — and picks up as soon as multica version answers. It won't install it for you: the options differ enough that guessing is worse than waiting
Up to date?	tells you what's newer and the line to run — on a self-hosted server the update is yours, and the server goes first
Signed in?	multica setup cloud, or multica setup self-host --server-url … for your own server
A workspace?	creates one, or confirms which if you have several
Daemon running?	multica daemon start — this is what actually executes the work
A runtime?	your agent CLI, auto-detected on PATH — install one if none, then restart the daemon

Installing or updating anything on your machine is reported, never done behind your back — it's your call, and Mops names exactly what each step does and why. One caveat worth knowing: daemon status answers about the CLI's own profile, so it can read stopped while the desktop app keeps runtimes online. multica runtime list is the probe that tells you whether work can actually run.

3 · Answer three questions, then it takes over

Once the environment is ready, Mops reads what you have and routes you — you never pick a command:

What you have	Where it goes
nothing yet	/multica-ops:init — it shapes the work with you, then staffs a team
a Multica workspace already	/multica-ops:join — audits it, then fixes in batches you approve
a backlog in Linear / Jira / a CSV	/multica-ops:import — moves it over, unassigned, nothing runs by itself
a question, nothing to build	/multica-ops:consult — an answer, and no machinery unless it leads there
a list of tasks, you'll set the order	/multica-ops:mops crew — executors and gates, you're the manager
a one-hour job	/multica-ops:quick <task> — one or two agents, build → review, no machinery

These are the commands you type — in full, /multica-ops:init, /multica-ops:join, so you copy rather than translate. Plain language always works too ("set up a team"). There is no bare /init and no short /mops: outside Claude Code there are no slash commands at all.

The interview adapts to you. Tired of questions? Say "you decide" and Mops proposes the whole setup as one list to confirm or tweak. Want to be closely involved, or barely? That is the first thing it asks — and it decides how much else you are asked at all.

4 · Getting the new version

Say "is there a new version?" or run /multica-ops:upgrade. Mops fetches the update itself. New content applies the next time a file is read — you only restart when the release adds or renames a command or touches a hook, and Mops tells you which case it is. Then it migrates your workspace, re-screens any imported skills, tells you the line to update the CLI (yours to run, and only when the team is idle), and finishes with a plain-language tour of what changed.

Upgrading from an older version? Slash commands exist only on a Claude Code plugin install, and older releases used different names — so /multica-ops:upgrade may say unknown command. Two paths that work from any version:

1. Say "upgrade" in plain language — no slash needed; Mops runs the flow and gives you the one line to run.
2. Update the install yourself, then restart your agent: claude plugin update multica-ops (Claude Code plugin) or npx skills add jamillazarev/multica-ops (skills.sh install).
3. The resident copy inside a Multica workspace has its own line — multica skill refresh <id>, which re-downloads from the source it was imported from and keeps the id and the agent assignments. Before it existed the only route was a second skill import, which mints a new id and leaves every agent pointing at the old copy. New in CLI 0.4.26; surface verified 2026-08-15, the behaviour not run against a live workspace.

And "updated everywhere" is a claim about a generated list, never about memory — run bash scripts/find-installs.sh before updating and again after. It walks every install of this skill on the machine, reports its version and its update route, and exits non-zero on a broken symlink or a silently stale copy.

Make the skill load, at both ends of the model range

Add a trigger rule to the always-on surface your runtime reads — open the multica-ops skill before acting on requests about running the company, its issues, agents, squads, budgets or shipping, on anything that spends, ships, deletes or changes the shape of the team, and on questions about how to run work.

Which surface, per runtime — this repository already ships two of them:

Runtime	The surface
Antigravity	rules/multica-ops.md, shipped here and always on
Gemini CLI	GEMINI.md, shipped here
Claude Code / Codex in a repo	that repo's CLAUDE.md or AGENTS.md
the Multica workspace itself	the shared guide and the Mops agent's instructions — there is no AGENTS.md in a workspace; the workspace is the company, not a git checkout

Both ends need it, for opposite reasons. A light model may not open the skill because it does not connect the request to it. A strong one may not open it because it does not need to: measured 2026-08-01 in the sibling project, three runs of "I want to build a macOS app that fixes system audio. Set it up." on a tier above the advisor's floor invoked no skill and read no corpus file — they wrote the app and compiled it. That is a defensible reading of the request, which is exactly why the rule cannot be left to discovery. Capability suppresses recourse to a methodology, and the anchor is what makes the choice explicit instead of implicit.