For small studios that use Claude Code
The back office we run our studio on, wired into Claude Code.
Muster is for small studios of 2 to 10 people who already use Claude Code.
Claude Code forgets everything when a session ends. Muster keeps the records that need to last in a Directus instance you host, and its plugin gives your agents access to them over MCP. When an agent finishes a task, it updates the same row you look at. Start with the read‑only demo.
Install
Agents update the records as they work.
Most back offices go stale because someone has to update them by hand. In Muster, the agents doing the work write the updates.
# in Claude Code /plugin marketplace add AnalogElk/muster /plugin install muster@muster
Add the marketplace first. /plugin install muster on its own fails until you do. Tested from a clean install of github.com/AnalogElk/muster on Claude Code 2.1.211, 2026‑07‑15.
# needs your board URL and a Directus API token /muster:connect https://your-board.example.com
The plugin adds four slash commands, one skill, one SessionStart hook, and one MCP server. .claude-plugin/plugin.json lists each one.
The first time an agent calls a Muster tool, Claude Code asks you to approve it. You approve once. Connecting needs a board URL and a token for a board you run. The public demo is a static snapshot and cannot be connected to.
Demo
Open the real system.
Both views are public and need no sign-in.
Team view
app.musterr.dev
Delivery, money, studio operations, and the analytics over all of it.
Client view
app.musterr.dev/client-portal
What a client sees: their projects, tasks, invoices, brand assets, and messages. Nothing else.
Once you are connected, /muster:board shows the same task rows from inside Claude Code, without opening the portal.
The demo is a static snapshot of a real deployment, captured on 2026‑08‑02, with synthetic data. Nothing you click is saved. It shows what the records look like; it does not show agent writes, and a clean install will not match it exactly.
What it stores
Eight areas, one database.
38 os_* collections, grouped by job. The schema snapshot has 50 entries in total, counting 6 support tables and 6 sidebar groups.
Work and tasks
Projects, tasks, sprints, and repositories. Each task row shows who claimed it, what blocks it, and whether it shipped, and it stays after the session that wrote it ends.
Demo: open Tasks. /muster:board shows the same rows from inside Claude Code.
AI cost ledger
A ledger of what the agents spent.
Demo: the AI ledger, synthetic data. Partly in Muster. See the features list.
Releases
A release record for each project, with a changelog.
Listed as in Muster on the features list.
Knowledge
A local retrieval service indexes your records and answers questions over them. The base snapshot has no knowledge-base or asset-library collections.
Retrieval runs locally. Answer generation uses the model provider you configure.
Client portal
Clients sign in to the same records you work in, limited to their own projects, invoices, brand assets, and messages.
Demo: open the client view, then compare it with the team view.
Money desk
Built: deals, proposals, invoices, payments, and expenses. It is a ledger, without double entry, bank feeds, or tax workflow.
Demo: Invoices and the revenue view. Synthetic data, read-only.
Studio operations
Organizations, contacts, activities, support, messages, and your own inbox. No payroll, hiring, or time tracking.
Demo: follow one organization through to its contacts and its activity trail.
Operations view
One dashboard for site, portfolio, delivery, and infrastructure analytics.
Demo: the dashboard, synthetic data. Connected-service views need your own accounts. Without them a view can be empty or explicitly mocked.
Built on the same records
Photo: private albums for clients
Private albums, magic links, favorites, zip downloads, and tips. Read the case study.
Internal only, not in the release.
Self-host
Run it on your own server.
Muster is a Docker Compose stack under the MIT licence. Your data stays in your own Postgres and Directus. There is no hosted plan.
# from inside Claude Code, once the plugin is installed
/muster:up
/muster:doctor
/muster:up starts the stack and /muster:doctor reports whether each part is healthy. The stack is defined in compose/compose.yaml and bin/elk-os, under the MIT licence.
Docker is the hard dependency. Before the first up, edit .env and set ELK_OS_WITH_PORTAL=false. The default is true, and up stops without a portal source. A fresh clone gets the Directus record layer: Postgres, Directus, the schema, and the seed data. The portal image elk-os-portal is private, so a clone cannot pull it, and building it needs a local checkout of analog-elk-front-end, which is not public. The portal is shown in the live demo and is not yet public. A clean self‑host does not reproduce every section of the demo.
Limits
What Muster does not do.
The first four affect whether you should install it.
- Muster does not make Claude smarter or run agents in the background. It gives instructed agents shared records and a protocol. Reach for it when the work outlives a session, or when more than one agent is running at once.
- The first MCP tool call requires approval in Claude Code. There is no zero-click connection. Approve it once.
- The public demo is synthetic and read-only. It proves shared reads, not agent writes or self-hosting. To watch an agent write a row, stand up your own.
- A clean self-host does not reproduce every demo section. The base snapshot omits demo enrichment and some portal dependencies. The portal image is private and is not yet public.
Also worth knowing
- The Money desk is not accounting software. There is no double-entry ledger, bank feed, reconciliation, tax workflow, or bookkeeping automation.
- Studio operations is not HR software. There is no payroll, hiring, benefits, leave management, or time tracking.
- Connected-service views require their own credentials. Without them, a view may be empty, unavailable, or explicitly mocked.
- Task claims are not atomic locks. Avoiding collisions requires disjoint assignments and worktree isolation. The row records the intent; it does not enforce it.
- Retrieval and embeddings can run locally. Answer generation uses the configured model provider, so nothing here claims that all assistant data stays on your box.
- Muster is unnecessary for work that fits in one context and has no back-office records to maintain. A single agent on a single small task does not need any of this.
The plugin added about 220 tokens per session when measured from a clean install on Claude Code 2.1.211 on 2026‑07‑15. Your numbers may differ. Any verification agents you run cost tokens on top of that, whether or not they find a problem.
More reading
How Muster was built.
Background, if you want it. You do not need either to install the plugin.
How Muster was built → Read the whitepaper → GitHub ↗
Both describe the original 2026‑06‑29 build and are kept as written, with dated notes where the product has since changed. The build log covers the agents and the bugs they caught. The whitepaper covers the same build in long form, with citations.