OpenClaw Fleet — Multi-Agent Skeleton

A sanitized, reusable skeleton for a self-hosted fleet of specialized LLM agents built on OpenClaw. It's both a reference architecture for personal multi-agent systems and a bootstrap template you can clone to stand up a fresh fleet.

🔒 Sanitized template. This repo contains structure only — no memories, no credentials, no personal data, no server details. Every persona, channel, and group here is a generic placeholder. The live system this is modeled on keeps all of that in memory/, credentials/, and sessions/ directories that are never committed (see .gitignore).

---

What this is

A single AWS host runs an OpenClaw gateway that supervises a roster of specialized agents. Each agent:

• owns a domain (orchestration, finance, projects, health, career, household ops, …),
• has its own persona (IDENTITY.md + SOUL.md), model, and Slack channel,
• reads/writes only the shared groups it's scoped to (a simple file-based access-control model),
• can run on a cron schedule (e.g. a morning brief) or purely on-demand,
• routes to any model through OpenRouter, and calls tools over MCP.

Architecture

flowchart TB
    subgraph Channels[Slack channels + DMs]
      C1[#general] & C2[#finance] & C3[#projects] & C4[#health]
    end
    Channels <--> GW

    subgraph Host[AWS host]
      GW[OpenClaw Gateway<br/>sessions · routing · cron · events]
      GW --> O[Orchestrator agent]
      GW --> A1[Domain agent A]
      GW --> A2[Domain agent B]
      GW --> A3[Domain agent C]
      O -. reads summaries .-> A1 & A2 & A3
    end

    GW --> OR[(OpenRouter<br/>any model)]
    GW --> MCP[(MCP tools)]

    subgraph Shared[Shared filesystem]
      SG1[(shared/<br/>team-wide)]
      SG2[(shared-groups/*<br/>scoped per domain)]
    end
    A1 & A2 & A3 --> SG2
    O --> SG1

Loading

The agent roster (archetypes)

Real fleets mix one orchestrator with several domain specialists. This skeleton ships generic archetypes you rename and reskin:

Archetype	Role	Schedule
orchestrator	Compiles a daily brief by reading every domain's summary; cross-agent coordination	cron (daily)
finance-analyst	Owns the finance shared-group; budgets, tracking, summaries	on-demand
project-manager	Tracks side-projects and ventures; status rollups	on-demand
_TEMPLATE	Empty agent — copy it to add a new specialist	—

Repo layout

openclaw.example.json     # sanitized gateway/agent config (placeholders)
REBUILD.md                # bootstrap a fresh fleet from this skeleton
docs/
  ARCHITECTURE.md         # how the pieces fit (gateway, agents, groups, cron, MCP)
  SHARED-GROUPS.md        # the file-based access-control model
agents/
  _TEMPLATE/              # copy this to add an agent
  orchestrator/           # example: the daily-brief orchestrator
  finance-analyst/        # example: a scoped domain specialist
  project-manager/        # example: a second domain specialist

Each agent directory holds: IDENTITY.md (one-liner card), SOUL.md (persona + scope), AGENTS.md (operating instructions + data access), HEARTBEAT.md (periodic tasks), TOOLS.md (environment notes).

Why it's interesting

• Separation of concerns across agents — each agent is small, single-purpose, and independently testable, communicating through well-defined shared files rather than shared memory.
• Capability-scoped access — a finance agent literally cannot read the health group; access is enforced by which directories each agent's instructions grant.
• Model-agnostic — every agent picks its own model via OpenRouter, so cost and capability are tuned per role.
• Disaster recovery — because the whole fleet is declarative (config + markdown personas), this skeleton rebuilds a working system from scratch; only memories and credentials need restoring.

See docs/ARCHITECTURE.md for the deep dive and REBUILD.md to stand one up.