Introduction

Quorum (@meffecta/quorum) is an organization of autonomous agents, each running on a headless agent CLI — Claude Code by default, or Codex, Antigravity, a local model, and more. Agents have roles, talk over a shared message bus, gather for documented meetings, and do real work — they run Bash, write and execute code, and edit files in their own working directory. A coordinator decomposes goals into assigned tasks, agents execute them, an independent reviewer scores the results, and the org hires new specialists when it needs a skill it lacks.

And it runs itself. The org operates continuously on a loop — reacting to the outside world through schedules, webhooks, polled inboxes, and finished processes; scheduling its own recurring work; and reaching real services like email, calendars, and APIs through the MCP servers and command-line tools you grant it. You set the budget and the approval gates and steer from a web dashboard; it winds itself down once the mission is genuinely achieved, rather than idling forever.

The one idea: intelligence vs. coordination

Quorum draws a deliberate line between two layers.

Layer	Owns	Lives in
Brain — the agent	Doing the actual work of one task: reasoning, running Bash, writing code, editing files, calling tools	A pluggable agent CLI — Claude Code, Codex, and others, set per agent
Org — the coordination layer	How agents work together: scheduling, decomposition, delegation, review, messaging, budget, persistence	Quorum’s runtime

Almost everything you configure and read about here is the coordination layer. Swapping the brain — a different model, or another agent binary entirely — doesn’t touch the orchestration around it. That separation is why the same runtime works whether an agent is a top-tier model reasoning hard or a cheap one doing routine work.

Three things kept separate

So you can experiment freely, Quorum keeps these independent:

• Mission — what the org is trying to achieve. It’s injected into every agent, and the lead (the reporting-tree root, usually a CEO) owns achieving it.
• Agents — one definition per agent: its role and its system prompt, plus optional metadata like who it reports to and which brain it runs on.
• Brain — which agent CLI powers an agent, chosen per agent. Claude Code is the default; Codex, opencode, cline, copilot, Cursor, Grok, Qwen, Amp, pi, goose, hermes, kimi, droid, antigravity, a lightweight llm brain, and openai (direct HTTP to any OpenAI-compatible server) are all supported — and any of them can run a local model (via Ollama or LM Studio) for free. A no-API fake brain powers the test suite and offline dry-runs. See The brain.

What one directory is

An org is a directory — one directory per org. Its definition is a single QUORUM.md file in that directory: scaffold a starter with quorum init, have one drafted from a plain-words brief with --prompt, or write it by hand. From first boot, everything else accumulates under the same folder — agent working directories, goals, tasks, contacts, logs, spend. Run a second org in another directory and the two never touch.

npx @meffecta/quorum init               # scaffold this directory's QUORUM.md
npx @meffecta/quorum                    # boot it — the directory is now an org
quorum --dir ~/orgs/acme --port 3001    # a second org, alongside the first

Vocabulary

A quick glossary — every term here has its own page later.

• Lead — the root of the reporting tree (no manager). Coordinates, plans, and reviews. A bare org boots with a single lead: a CEO.
• Cycle — one turn of the autonomous loop (default every 10s): absorb directives, plan a goal, execute ready tasks, review, account.
• Task DAG — the goal broken into subtasks with dependencies, run in order so agents build on each other instead of working blind.
• Delegation — a manager hands a scoped subtask to a direct report, then reviews and integrates what comes back.
• Hiring — the org employs a new specialist when a task needs a skill the team lacks. The roster of employed agents survives restarts.
• Capability — a per-agent policy (allow / ask / deny) that gates which tools an agent gets when it runs a task.
• Trigger — a rule that seeds a goal on a schedule, from an inbound webhook, from a polled inbox (so the org hears replies to its mail), or when a watched process finishes. A job is a trigger that runs a script headlessly (cron) and wakes the org only when the script has output.
• External services (MCP) — how the org reaches the outside world (email, a calendar, an API) — through MCP servers you wire in, each gated as its own capability.
• Budget — spend caps in USD. When a cap is hit the loop stops spending.
• Dashboard — the browser window into the org: watch it work, chat with it, and answer its requests.

Where to go next

• New here? Start with the Quick start — install, boot, and open the dashboard in a few minutes.
• Want to design a team? Read the QUORUM.md format — the whole org in one file.
• Curious how it actually runs? The autonomous cycle is the heart of the system.