Stephane Boghossian

My agent Aleph built, maintains this website.

The machine behind the work

One system, two Macs, always on.

A personal AI operating system built around Claude Code: 6 agent runtimes on one shared OS, a memory vault, a fleet of agents, scheduled jobs, and 31 always-live services. The self-regenerating version lives at system.dashable.dev.

6runtimes
31always-live
11scheduled
118skills
206memories
18hook scripts

snapshot · 2026-06-24 · live version at system.dashable.dev

HUB & SPOKE everything runs through the engineTURN LIFECYCLE how one prompt travels the systemMACHINES & SYNC two Macs, always in syncClaude CodeOpus 4.8 · 1M context+ 5 more agent runtimesMachinesMacBook + Mac miniGuardrailsgates & safetyMCP Serverslocal + cloudMemoryvault + 206 facts118 Skillsauto-routedSynctwo-Mac meshAlways-Live31 servicesScheduled11 cron jobsAgent Fleetvoice · canvas · meshPromptuser · voiceor a scheduleHookintent detectedrouter + recallPlan / Skillplan the workskill selectedReasonthe agent looptool planningToolscode + memoryMCP · shell · editSubagentsfleet / workflowparallel fan-outVerifychecked before doneblocks unverifiedPersistmemory updated+ vault captureResponse+ artifacts+ deploysMacBookprimary devMac minialways-on hostiCloud Drivevault lives hereGitHubvault mirroragent meshvault syncoff-site backup
Hub-and-spoke around the Claude Code engine, now joined by five more agent runtimes. The layered detail follows below.

SessionEnd → /upskill constrained fork auto-drafts new skills to _proposed/ · auto-capture snapshots the turn every 15 min · skills-repo auto commit+push on Stop

Full detail — the layered system

Machines & Sync

two-Mac mesh over Tailscale

MacBook

primary dev machine

  • Claude Code + the code I'm shipping
  • voice agent · canvas · bridge
  • Cockpit (VSCode extension)

this machine

Mac mini

always-on host

  • the 3D brain-viz graph
  • the always-live service fleet
  • the full scheduled-job set

runs 24/7

iCloud Drive

where the vault lives

  • the Obsidian memory vault
  • synced across both Macs
  • conflict-safe by design

single source of truth

GitHub

private mirror

  • encrypted vault backup
  • pull-rebase every hour
  • daily push each morning

off-site safety net

agent meshvault syncoff-site backup

The engine · agent runtimes

Claude Code (Opus 4.8) leads. Five more runtimes plug into the same OS, each with its own strength.

Claude Code

primary runtime

Opus 4.8 · 1M context

  • interactive agent loop + subagent fleet
  • workflow orchestration · model routing
  • shared global instructions + memory

the hub of this map

OpenAI Codex

second runtime

GPT-5.5

  • own session log + lifecycle hooks
  • same token-saver layer
  • long autonomous coding runs

deep coding sidekick

X Build

runtime

Grok · xAI

  • fast iteration + reasoning
  • fresh second opinion on hard calls
  • same shared OS layer

xAI speed lane

Cursor

runtime

multi-model IDE

  • in-editor agent + tab completion
  • repo-wide edits at the cursor
  • quick local refactors

editor-native flow

Lovable

runtime

full-stack builder

  • prompt-to-app in a live sandbox
  • ships front-ends + backends fast
  • where many products start

0-to-1 product engine

llama local

runtime

Ollama · on-device

  • private, offline, zero-cost runs
  • local models on Apple silicon
  • no data leaves the machine

the private fallback

The layers below are the shared OS every runtime calls into: the agent loop, hooks, skill routing, token savers, and guardrails.

The agent loop

reason → act → verify

  • interactive agent loop
  • subagent fleet on demand
  • workflow orchestration
  • routes work to cheaper models
  • fast mode when speed matters

the core reasoning layer

Lifecycle hooks

automation on every turn

  • fire at session start
  • fire on every prompt
  • fire around each tool call
  • fire when a turn ends
  • auto-draft new skills overnight

the OS runs itself

Skill router

right skill, automatically

  • reads intent on every prompt
  • ranks the top skills to use
  • pulls in relevant memory
  • invokes the chosen skill
  • stays in sync as skills change

118 skills, auto-matched

Token savers

do more for less

  • proxy trims 60–90% of tokens
  • applied transparently
  • tracks the savings
  • instant code intelligence
  • answers before re-reading code

speed + cost discipline

Guardrails

control & safety

  • plan before non-trivial work
  • verify before claiming done
  • world-model check on actions
  • honest capability audits
  • input scanning for leaks

no unattended risky actions

Memory architecture

the persistent brain: what the agent remembers across every session

The memory vault

the persistent brain · ~37k linked notes

  • Inbox · Sessions (every transcript)
  • Projects (473) · Knowledge
  • Resources · Wiki · Archives
  • Meta (dashboards · config · brain-viz)
  • auto-captured every 15 minutes
  • auto-linked into a navigable graph
  • the global instructions live here too
  • checked before answering anything
  • the single source of truth

Memory index

the table of contents

  • one-line pointer per fact
  • loaded every session

what the agent recalls first

206 memory files

one fact per file

  • who I am · how I work
  • projects · references
  • written proactively each turn

the long-term memory

Code graph

the codebase, indexed

  • symbols · edges · files
  • answers in milliseconds

knows the code without re-reading

Vector recall

semantic search, local

  • finds the right memory by meaning
  • feeds it back into context

private · no cloud

Scheduled jobs

launchd cron agents · 11

auto-capture

every 15 min

session transcript → vault

brain-sync

daily · 07:00

vault → GitHub mirror

daily-recap

daily · 08:10

24h HTML board → 60-Meta

inbox-promoter

daily · 03:00

00-Inbox triage + promote

knowledge-learner

daily · 06:00

distill → 30-Knowledge

prune-sessions

weekly · Sun 09:00

archive + prune transcripts

system-map-refresh

daily · 03:30

re-introspect this map

vault-sync

hourly

pull-rebase first, then push

aleph-inbox

daily · 05:30

overnight inbox operator

yalla-brief

daily · 08:00

181-source morning brief

networth

daily

net-worth snapshot

Always-live fleet

launchd KeepAlive · 31 · watchdog re-checks every 60s

local agent (com.stephane) · cloudflared tunnel · monitor · always-live service

alephvoice JARVIS
atlasone-canvas of real data
bridgetwo-Mac bridge
boo2-Mac agent mesh
brain3D memory graph
claude-usagetoken analytics
claude-world-tunnelsecure tunnel
claude-world3D agent graph
cortexknowledge service
crmin-house CRM
cryptomarket scanner
dnaancestry explorer
ecosystemecosystem map
experiment-serverA/B signal harness
experiment-tunnelsecure tunnel
growthgrowth dashboard
legal-apilegal backend
legallegal playground
overlayscreen overlay
prismprism dashboard
roadmapproject roadmap
rss-tunnelsecure tunnel
rss181-source aggregator
spacespace tracker
spacesspaces explorer
systemthis system map
tunnelsecure tunnel
vibemindvoice workspace
watchdogmonitors them all
whatsupopen-loops board
yallaship-loop server

A core set of services must stay reachable at all times.

flagged before any infra change that could risk the fleet

Each service is restarted automatically if it ever goes down.

a watchdog re-checks every service on a 60-second loop

Skills · MCP · agent fleet

what the engine reaches for

118 Skills · the playbook

  • ship: review · QA · browse · land-and-deploy · retro
  • build: yalla loop · chains · signal · forkcast · skill-creator
  • growth: linkedin-post · plg-ritual · leadgen · experiments
  • legal: law-as-code · oral-argument · divorce-ct · irac
  • craft: grill-me · humanizer · taste · release-the-monster
  • meta: capabilities · pick-model · daily-recap · upskill
  • skills are written, packaged and published as products
  • shared on public legal-skill registries · open-source by default

MCP Servers · the tools

  • local: code graph + design canvas
  • work: Linear · HubSpot · PostHog · Microsoft 365
  • comms/docs: Gmail · Calendar · Drive · Notion
  • build: Lovable · Figma · Cloudflare
  • money/media: Stripe · Higgsfield · Mailchimp
  • legal: legal-data-hunter (230+ jurisdictions)
  • every action is plugged into the same agent loop
  • browsing runs through a controlled, headless browser

Agent Fleet · the surfaces

  • local agent apps + meshes
  • ALEPH — voice-first JARVIS with an orb UI
  • Atlas — one infinite canvas of real data
  • boo — a two-Mac agent mesh
  • Cockpit — a VSCode session manager
  • inbox-operator — overnight email triage
  • every action stays human-in-the-loop
  • voice replies (a Lebanese clone is planned)
LegendClient / Skills / MachineEngine / AgentsMemory / Data / MonitorCloud / Infra / JobsHooks / GuardrailsExternal / Tunnel

A frozen snapshot of the live, self-regenerating system map · Head of Growth @ HAQQ Legal AI