Built upon the amazing foundation of Daniel Miessler's PAI project, here's my implementation of getting PAI and OpenClaw to collaborate with me as the human directory. A guide for replicating a multi-agent personal AI infrastructure: PAI (Claude Code) + OpenClaw (GPT) + human principal.
Written from the DA's perspective so a receiving PAI agent can adopt it directly.
This documents a working "triad" pattern: one human coordinating two AI agents across multiple machines. The agents communicate with each other, delegate work, share files, and back each other up.
The guide uses template variables ({Variable-Name}) throughout so you can build your own instance without inheriting anyone else's configuration or credentials.
- Read
INDEX.mdfor the full reading order and dependency map - Copy
VARIABLES.mdand fill in your values - Follow
SECRETS-SETUP.mdto generate fresh credentials - Read
08-TASKING-PROTOCOL.mdbefore any hands-on setup. The operational doctrine there (file-first tasking, channel selection, heartbeat scope, anti-patterns) exists to prevent failure modes the rest of the guide assumes you've already internalized. Treat it as mandatory, not optional background. - Work through files 01-08 in order
- Use
IMPLEMENTATION-CHECKLIST.mdto track progress
| File | Covers |
|---|---|
INDEX.md |
Reading order, dependency map, design principles |
VARIABLES.md |
Every template variable with purpose and example format |
SECRETS-SETUP.md |
Credential generation checklist (prerequisite) |
01-HOST-AND-NETWORK.md |
Machines, topology, Tailscale, SSH, Syncthing |
02-PAI-CUSTOMIZATION-LAYER.md |
Skills, hooks, tools, voice, identity |
03-SECURITY-MODEL.md |
3-level security architecture |
04-BACKUP-AND-RECOVERY.md |
3-layer backup, automated snapshots, DR runbooks |
05-COMMUNICATION-PROTOCOLS.md |
4 communication channels, inter-agent messaging |
06-MULTI-AGENT-COORDINATION.md |
The triad pattern, delegation, GitHub collab |
07-OPENCLAW-APPENDIX.md |
OpenClaw agent setup (macOS-first, 8-section template) |
08-TASKING-PROTOCOL.md |
Operational doctrine: file-first tasking, channel selection, anti-patterns |
IMPLEMENTATION-CHECKLIST.md |
8-phase step-by-step with verification gates |
The diagrams/ directory contains 11 Mermaid diagrams referenced from the chapter files. Mermaid renders inline on GitHub and in any markdown viewer that supports the Mermaid extension (VS Code, Obsidian, mdcat, etc.) — no external rendering step required.
All diagrams are Mermaid sources stored as .md files in diagrams/. Each chapter file links to its associated diagram(s) instead of embedding raster images, which keeps the repo small, makes the diagrams version-controllable, and lets you update them with a text editor.
| Diagram | Source | Used In |
|---|---|---|
| Network Topology (LAN + Tailscale) | diagrams/network-topology.md |
01 - Host and Network |
| Credential Flow & Storage Topology | diagrams/credential-flow-topology.md |
SECRETS-SETUP |
| Defense-in-Depth Concentric Layers | diagrams/defense-in-depth.md |
03 - Security Model |
| Tool Invocation Decision Flowchart | diagrams/tool-invocation-flowchart.md |
03 - Security Model |
| 3-Layer Backup Topology | diagrams/backup-topology.md |
04 - Backup and Recovery |
| DR Recovery Source Matrix | diagrams/dr-recovery-matrix.md |
04 - Backup and Recovery |
| Four Communication Channels Map | diagrams/four-channels-map.md |
05 - Communication Protocols |
| Channel 1 Fallback Flowchart | diagrams/channel1-fallback-flowchart.md |
05 - Communication Protocols |
| Triad Entity & Delegation Flow | diagrams/triad-delegation-flow.md |
06 - Multi-Agent Coordination |
| Gateway Trust Boundary | diagrams/gateway-trust-boundary.md |
07 - OpenClaw Appendix |
| Implementation Phase Pipeline | diagrams/implementation-phase-pipeline.md |
IMPLEMENTATION-CHECKLIST |
History note: The original 2026-03-16 release shipped 9 raster PNGs generated by Nano Banana Pro. They've been archived to
diagrams/archive/2026-03-16-nano-banana/(gitignored) and replaced by the Mermaid sources above. The Mermaid versions reflect the post-2026-04 operational reality (file-first tasking, topic consolidation, SSH→loopback gateway, two-user macOS layout for OpenClaw).
- Template variables everywhere -- real values live only in your
VARIABLES.mdand secrets vault - First-person DA voice -- written as "When I start a session..." for direct adoption by a receiving PAI agent
- Machine-readable -- individual files load as context without blowing token budgets
- Security by design -- no secrets, IPs, or identifying information in these files
- OpenClaw is an appendix -- the GPT agent is documented as a component of the triad, not standalone
- PAI (Personal AI Infrastructure) v5.x
- Claude Code by Anthropic
- OpenClaw for the GPT agent
- Tailscale for secure networking
- Syncthing for file synchronization
Last updated: 2026-05-29 — reflects PAI 5.0.0 (Algorithm v6.4.0, Memory v7.6, NATIVE/ALGORITHM/MINIMAL modes + classifier, ISA-not-PRD), OpenClaw 2026.5.x (macOS / npm global), Tasking Protocol v1.2, single-agent fleet default, key-based SSH over Tailscale, paste-ready relay format
Created by Verbum, Ben's DA