System

Umbrella

A layer on top of Claude Code that runs every project, agent, and automation from one place.

The layer

Everything Claude Code does, organized into one system.

Claude Code already brings agents, skills, memory, and tools. Umbrella is the layer that sits on top and makes them work together. One version-controlled folder holds every area of work and life, each as its own organization with its own rules, knowledge, and agents, all running on shared infrastructure. No scattered one-off chats. Every session boots into a context that already knows who you are and what it's allowed to do.

Your workprojects · deliverables
appssitesdocs
Umbrellathe layer
centralizestandardizeautomate
Claude Codefoundation
agentsskillsmemoryMCPhooks
01Organizations

Separate worlds that never bleed together.

Work is split into separate organizations. Each is self-contained, with its own knowledge, rules, data, and agents, and none can reach into another. Each org sets its own boundaries, down to what it ingests and what its agent can touch. The boundaries are the point: context stays where it belongs.

Org A
ORG.md
knowledge
tasks
ingest
agents/
Org B
ORG.md
knowledge
tasks
ingest
agents/
Org C
ORG.md
knowledge
tasks
ingest
agents/
Shared core

All orgs sit on the shared core, with a separate maintenance agent that tends the structure itself.

02Agents

Every org runs its own agent.

An agent is a Claude Code session launched inside an org's folder. It loads the full picture at startup: the standing brief it follows, the org knowledge, the guardrails, the open tasks, and the memory it carries. It writes a session journal as it goes, so the next session picks up exactly where the last one left off. An org can hold more than one.

CLAUDE.mdknowledgeguardrailstasksmemory
Agent
session logcommits
03Homebase

A homebase, not a container.

The umbrella doesn't hold the projects. It holds the knowledge, skills, tools, and rules, plus a registry of where each project lives and what the agent is allowed to touch. The agent runs from here and reaches out to projects in their own repos across the machine. The finished work lands there. The umbrella stays the place it all runs from.

Umbrella
knowledgeskillstoolsrulesregistry
reaches into
App repo~/Developer/…
Website repo~/Sites/…
Docs & notes~/Documents/…
04Structure

One layout, everywhere.

Every org follows the same shape, so any agent knows where to look without being told. Knowledge, tasks, ingested data, and shared tools always live in the same place. That predictability is what lets the whole thing scale: a new org is stamped from a template in one pass.

umbrella/
umbrella/                       # one private git repo
├── _shared/                    # shared toolbox, used by every org
│   ├── bin/                    # scripts that wire it together
│   ├── lib/                    # the data-sync engines
│   ├── skills/                 # skills shared across orgs
│   └── templates/              # blueprint a new org is stamped from
├── agents/maintenance/         # caretaker agent for the structure
└── orgs/
    ├── org-a/                  # ┐
    ├── org-b/                  # ├ each org, same shape
    └── org-c/                  # ┘
        ├── ORG.md              # the org's knowledge, always loaded
        ├── projects.yaml       # registry: what to touch, where it lives
        ├── GUARDRAILS.md       # the rules: what needs approval
        ├── ingest/             # incoming data, as dated markdown
        ├── knowledge/          # durable reference
        ├── tasks/              # actionable items, one file each
        ├── skills/             # skills scoped to this org
        └── agents/main/        # the agent you launch
            ├── CLAUDE.md        # standing brief, read every session
            ├── memory/         # typed facts it remembers
            └── sessions/       # a dated journal of its work
05Shared core

Write it once, every org gets it.

Common scripts, sync engines, skills, and templates live in one place and feed every org. Improve something at the core and the improvement reaches every agent at once, with no copy-and-paste drift. The templates are the blueprint a brand-new org is stamped from.

_shared
Org A
Org B
Org C
06Skills

Reusable abilities, authored once.

Skills are packaged capabilities an agent can pick up: a way of writing, a way of designing, a repeatable workflow. They're written once and linked into every agent automatically at session start, so a new skill is available everywhere the moment it exists.

house-styleApply a fixed format and tone.
data-triageTurn raw feeds into action items.
deep-researchRun multi-source research and synthesize it.

And a growing library, each one written once and shared across every org.

07Knowledge & memory

It remembers.

Nothing has to be re-explained. Each org builds a reference over time, and the agent carries typed memory from one session to the next.

Knowledge baseA reference each org builds up over time. The agent reads it at the start of every session, so it starts informed instead of blank.
MemoryTyped facts that persist across sessions, like preferences and past decisions, so the same ground is never covered twice.
08Ingestion

Real context, pulled in when it helps.

Small sync engines pull in email, calendar, chats, and meeting transcripts, and write them out as dated markdown the agent can search. Each org decides what it pulls in and how often: some on a schedule via launchd, some only on demand, some nothing at all.

emailcalendarchatstranscripts
.md
Agent
09Maintenance

The plumbing runs itself.

Skills wire into each agent at session start, a reminder fires before the working context would be trimmed, and launchd keeps the data feeds current. Everything is plain markdown in a private git repo, committed automatically, so the full history is there to read, revert, or audit at any time.

c808434auto: sync
8d90e54session: journal updated
a9bc97ctask: captured from triage
919b316skill: added to shared core
e72fe3aknowledge: reference updated
10Privacy

Boundaries that hold.

Scoped by designAn agent launches inside a single org and stays within it. Its brief and guardrails keep it scoped to that org and the shared toolbox; reaching into another is treated as out of scope.
You set the reachEach org decides what it ingests and what its agent can touch. Those limits are configured up front, per org, not bolted on after.
Approval doesn't carry overAnything irreversible or outward-facing waits for a yes, and a yes in one session never authorizes the same action later.
How it was built

Built entirely with Claude Code.

Umbrella is not a demo. It is the layer behind every other project on this site, used every day to do real work.

Claude CodeMCPGitlaunchdShellMarkdown