muster

Assemble AI agent teams with extreme character-driven personalities.

Install

curl -fsSL https://raw.githubusercontent.com/ddnet-repo/muster/master/install.sh | sh

Requires: python3 and opencode.

Usage

cd your-project
muster init       # wizard: name agents, pick characters, choose a lead
opencode          # open the project — your agents are ready

Other commands:

muster info       # show project setup and team roster
muster add        # add a team member
muster remove     # remove one
muster reset      # wipe everything and start over

What muster init does

Asks you:

• Project name
• What your agents should call you (Commander, Boss, My Lord, etc.)
• For each agent: codename, what they do, who they are (as a character), autonomy level
• Which agent is the lead (coordinates, delegates, doesn't code)
• Models per agent (opus for lead/thinking, sonnet for coding)

Then it generates:

.opencode/agents/
  <name>.md       <- OpenCode agent definition with extreme character personality
  ...

comms/
  team.json       <- team roster and config
  <name>/
    journal/      <- session logs in character voice (captain's log)
    notes/        <- persistent memory across sessions
  ...

opencode.json     <- OpenCode config
AGENTS.md         <- project instructions

How it works

Muster generates OpenCode agent definitions with extreme character personalities. Each agent is a fully realized character — not a "tone" or a "style," but a complete personality that never breaks.

Coordination happens through OpenCode's native agent teams system: message passing, task management, peer-to-peer communication, auto-wake. Muster doesn't manage any of that — OpenCode does.

What muster adds:

• Character layer — Agents ARE their characters. Darth Vader doesn't politely suggest improvements. Scooby-Doo literally talks like Scooby-Doo. The character dynamics drive the team dynamics.
• Persistent memory — Each agent has journal/ (session logs in their voice) and notes/ (working memory). Read the journals to see your project's history told by characters.
• Recipes — Pre-built teams. Mystery Inc., Wu-Tang Clan, Star Wars dark side, Golden Girls, and more.

Recipes

Pre-built team configurations in .muster/. Use them with muster init --from <file> or select interactively during init.

Recipe	Theme	Agents
atelier.json	Famous Artists	leo, frida, dali, basquiat, bob
breaking-bad.json	Breaking Bad	walt, jesse, gus, mike, saul
dark-side.json	Star Wars	palpatine, vader, tarkin, maul, thrawn
enterprise.json	Star Trek TOS	kirk, spock, bones, scotty, uhura
golden-girls.json	Golden Girls	dorothy, sophia, blanche, rose
hells-kitchen.json	Celebrity Chefs	ramsay, bourdain, child, white
magnum-pi.json	Magnum P.I.	magnum, higgins, rick, tc, robin
mighty-boosh.json	The Mighty Boosh	howard, vince, gregg, naboo, bollo
mystery-inc.json	Scooby-Doo	fred, velma, daphne, shaggy
rogues-gallery.json	Batman Villains	riddler, joker, bane, catwoman, freeze, scarecrow
star-trek-villains.json	Star Trek	khan, q, borg-queen, gul-dukat, chang
writers-room.json	Famous Authors	hemingway, poe, wilde, kafka, twain
wrestlemania.json	Pro Wrestling	undertaker, rock, stone-cold
wu-tang.json	Wu-Tang Clan	rza, gza, meth, deck, odb

Why characters?

When an LLM has a vivid character to inhabit, it commits harder to the role's constraints and communicates more distinctively. The character descriptions are deliberately extreme so that even after context dilution across a long session, the personality still comes through.

A backend engineer named "GZA" who approaches code like Liquid Swords — surgical, deliberate, no wasted bars — will produce more distinctive and consistent output than a generic "Backend Developer" prompt.

The friction between characters is intentional. If Vader force-chokes a subordinate for sloppy code, that's a feature. If Palpatine puts Vader in his place, that's the hierarchy working. Let them be who they are.