My Agent — Personal AI Collaborator Architecture
OpenClaw runtime · Tailscale-protected · model-agnostic. A layered guide to how a personal AI agent thinks, remembers, and acts.
Github · Repositories · Sync Layer
Durable sync between local, repos, and VPS
Agent Brain
private
Source of truth for
identity, memory, rules,
tools, and project context.
Agent Portal
public-facing
Approval pages, dashboards,
hosted outputs. Deploys via
Netlify to a public subdomain.
Skills Master
optional
Portable skills and MCP
server configs, kept in sync
across environments.
AI Model Layer · Priority Routing
Model-agnostic. No single-vendor lock-in.
1
OpenAI Codex
subscription · primary coding + reasoning
2
Anthropic Claude
subscription · long-context reasoning, drafting
3
Paid OpenRouter models
fallback layer · routes across multiple vendors on demand
Agent Brain · Configuration Files
Read at startup — shapes every session
IDENTITY.md
USER.md
MEMORY.md
TOOLS.md
HEARTBEAT.md
AGENTS.md
project files
reference files
Identity tells the agent who it is. User describes who it's working for.
Memory holds what it knows. Tools defines what it can do.
Heartbeat sets the recurring rhythm. Agents.md describes sub-agents.
Reference and project files give domain context.
Human · Local Environment
User & Inputs
U
User
the human in the loop
OpenClaw Web Portal
built-in browser chat with the agent
Gmail Inbox
agent reads + drafts on user's behalf
Google Drive Inbox
drop files here for the agent to ingest
Local Machine
user's laptop · code, files, Git client
GitHub Sync (from local)
push Brain edits, code, configs
Manual Brain Edits
optional — directly edit identity/memory
Public inputs (portal, Gmail, Drive) reach
the agent over the internet. The local
machine connects through the encrypted
Tailscale mesh.
The user never SSHs to a public admin
port. That surface is closed.
Tailscale
private
encrypted
mesh network
local ⇄ VPS
Vultr VPS · Ubuntu Host
OpenClaw Runtime
Self-hosted, always-on container for the agent
OpenClaw Container
Manages sessions, sub-agents, cron, browser, MCP connections.
My Agent
the configured personal orchestrator
• Reads Brain config at startup
• Routes model calls in priority order
• Calls connected tools and MCPs
• Writes memory + outputs
• Gates risky actions through approval
• Spawns sub-agents and cron jobs
My Agent ≠ the VPS. The agent is the configured runtime living inside
OpenClaw on the VPS. Swap the config, and you get a different agent.
OpenClaw provides the body (sessions, scheduling, browser control).
The Brain repo provides the mind (identity, memory, rules, tools).
The two together produce a stable, persistent personal AI collaborator.
Approval & Safety Loop
Required for public, spend, irreversible, or external actions
Agent drafts
the action
Discord
approval ping
Portal
approval page
User decides
(approve/edit/deny)
Log +
execute/abort
Connected Tools · APIs · MCPs
How the agent reaches the outside world
A · Meetings & Comms
Fathom MCP
meeting search, summaries, transcripts
Discord
notifications, conversational output
Twilio SMS
SMS backup for high-priority pings
Channels the agent uses to talk WITH the user.
B · Google Ecosystem
Google Workspace / CLI
Gmail · Calendar · Docs · Sheets · Contacts
Google Drive folders
Inbox · Processed · Reference · Projects · Scratch
User's primary work environment + the raw knowledge layer.
C · Models & Media APIs
OpenRouter API
OpenRouter Image
OpenAI API
xAI API
ElevenLabs
Cloudinary
Direct API access for image gen, speech, and hosted media.
D · Web · Data · Research
Apify
Firecrawl
Context7
Playwright
Scrape, crawl, monitor, run browser tests, look up
current SDK and API documentation.
E · Infra · Project
Firebase
Portal backend / project DB
Netlify
Portal deploys + site management
ClickUp API
task / project access
The infrastructure stack the agent operates on.
F · OpenClaw-native
Browser Control
Cron / Heartbeats
Sessions / Subagents
Nodes (paired devices)
Image / PDF / Media tools
automate the browser
scheduled + recurring jobs
child agents for background work
control paired devices
media analysis + generation
input
private Tailscale connection
sync
commit back
model call
reads identity
tool / API call
Outputs · User-Facing Surfaces
Where the agent's work shows up
Discord messages
notifications, approval pings, conversational replies
Email
to the user, plus confirmed outbound when acting on their behalf
Google Drive files
organized project outputs filed into the right folders
Agent Portal pages
approval pages, dashboards, deliverables on the public subdomain
GitHub commits / pushes
to Brain, Portal, or Skills Master when approved
Cloudinary-hosted media
images, files, signed URLs for shareable assets
Twilio SMS
high-priority text backup, breaks-glass channel
Browser artifacts / screenshots
visual proof for tasks that ran headlessly
Memory · Knowledge Flow
Four layers, most curated at the top
Layer 1 · Curated memory in Agent Brain repo
most distilled — pinned facts, rules, key relationships, decisions
Layer 2 · Daily notes
Memory/YYYY-MM-DD.md · rolling working log of the day
Layer 3 · Raw / archive — Google Drive
Inbox · Processed · Reference — ingested source material
Layer 4 · Created outputs — Google Drive
Projects · Scratch — what the agent produces, organized for retrieval
memory write
retrieval
output
memory write
retrieval
Optional Add-On
Garry Tan's gBrain — semantic memory upgrade layer
Works WITH the OpenClaw memory mechanism as an advanced semantic layer. It does not replace the default memory.
A long-term memory + knowledge-graph layer that ingests raw notes, decisions, documents, links, and project context. Sits
alongside the Brain repo and Google Drive memory layers, plugging into the agent to improve retrieval, continuity, and semantic recall.
Particularly useful when explaining ideas to a curious student: the agent can pull in the right precedent, snippet, or analogy on demand.
in progress
optional upgrade
not core runtime yet
complementary, not replacement
gBrain Layer
Semantic memory · knowledge graph
→ ingests notes, decisions, docs, links
→ adds vector + graph retrieval
→ improves continuity across sessions
→ student-friendly explanations
→ runs alongside Brain + Drive memory, not instead of
augments memory
semantic recall to agent
Legend · Arrow & Zone Key
input
user reaching the agent
sync
GitHub push / pull
private Tailscale
local ⇄ VPS only, encrypted
model call
agent ⇄ LLM via priority router
tool / API call
MCP or API request
output / approval
user-facing surface, gated
memory write
into Brain or Drive
retrieval
agent reading memory
optional gBrain
semantic upgrade flow