12 — Agent OS: the internal operating system

Prime Directive #3, made literal. The plan says “Use the tools internally first — the internal machine is the product being built” (01-business-plan.md, §0.3), and the thesis says the internal discipline is a standing intake engine that runs hunt → evaluate → categorize → master → deploy on the market so clients never have to (00-thesis.md). Agent OS is the cockpit that engine runs in. This section is how J&M maximizes it — as the daily driver behind the revenue engine, and as the reference build that makes the “operating system you own” pitch true because we live on it first (Node Zero).

STATUS: LIVE. Agent OS is deployed on the VPS at agentos.jacobdart.com (Cloudflare Tunnel + Access, only yakobdart@gmail.com admitted), bound 127.0.0.1:3737, wired to the box’s Hermes + Claude and the platform stack (Twenty CRM, Metabase, n8n, Nextcloud vault, Cognee). See the agent-os-agentic-os runbook + personal-ai-os/infra/agent-os/.

2026-07-06 update — the whole platform is now behind the same front door. The tunnel was extended to aios / crm / metabase / n8n .jacobdart.com (Access app aios-platform, same Only-Jake policy) — Twenty, Metabase, n8n, Nextcloud and the agent chat are reachable from any device, no SSH tunnel needed. Two standing cron automations run on the box’s Hermes (from the “15 Levels” playbook, sources/ hopper): a zero-token watchdog every 30 min (--no-agent, script-is-the-job, silent unless something is broken) and a daily 7am ops brief (pre-run script gathers platform/disk/front-door/vault data for free; the model only formats). Code: personal-ai-os/infra/cron/. Evaluated and rejected from that playbook: Hermes budget: caps (no such schema in our build — cost control stays at the OpenRouter key + free-model routing) and multi-profile splits (single profile + delegation is enough at current scale; revisit at 3+ Concierge clients). The profile-distribution idea (agent setup as an installable repo) is noted as the packaging pattern for the Company Brain offer — mom-setup already proves it.

Internal-only section (like 08/11): kept out of build_share.py PAGES, so it isn’t in the public plan and can name the internal machinery directly.

What it is (grounded, not hype)

A local Next.js “Mission Control” dashboard — a lobby, and each sidebar tab is a room with its own specialist model behind it (install/0-HOW-IT-ALL-WORKS.md). ~40 tabs, grouped into four categories:

Plus a goal taxonomy (~/.agentic-os/config.json → goalCategories): Business · Health · Personal · Learning · Side Project.

The honest product truth (from 30-DAY-ROADMAP.md, “keep what you use, ignore the rest”): nobody runs 40 tabs. Most people live in 3–4 daily, with the rest as deep modules on demand. J&M’s job is to pick its 3–4 and instrument them — not to boil the ocean.

The maximizing plays — Agent OS mapped to the revenue engine

Overlay of the plan’s own pipeline (07-automation-connector-map.md, Part 1: Find → Qualify → Pitch → Close → Deliver → Expand) onto the Agent OS tabs that drive each stage.

Reconcile: our prospecting pipeline vs. Agent OS Leads

We already built a specialized front end — pipeline.py (SEO-audit → email-draft → report) feeding Twenty CRM with an outreachStage lifecycle (see the prospecting-pipeline automation, ~/projects/ai-business-prospects/). Agent OS Leads/Outreach is the generalized one (find → enrich → score → send). They don’t fight:

• Our pipeline wins where the hook is the audit — SE-Idaho ICP, low-hanging SEO/GEO fixes, the Growth Plan report. That’s the wedge; keep it.
• Agent OS Leads wins for breadth — a new vertical/city where we don’t yet have a report engine, or for enrichment/scoring on a raw list.
• One system of record: Twenty. Whatever the front end, every prospect lands there from first contact (Cold), a deal Opportunity only at Replied/Booked. Agent OS fronts Twenty; it doesn’t replace it.

Orchestration: hand Hermes a standing goal

The highest-leverage tab is Hermes Goal Mode — long-running autonomous goals that survive restarts (--yolo, quiet, max-turn capped so it can’t run forever). Give it recurring chores: “scan this week’s tool launches, file the keepers into the hopper” (the thesis intake engine, automated), or “re-audit the 135 registry sites monthly, flag new gaps.” The 3-minute cron interrupt guard stays on. Model-routing rule: Hermes/Jarvis/video point at a strong model (real Claude / OpenRouter N2); local Ollama is only for the free on-device builder — anything else degrades output (0-HOW-IT-ALL-WORKS.md).

Dogfood → sell (why this is a revenue lever, not just tooling)

The site already promises to “turn your business into an operating system you own” (01-business-plan.md, §positioning). Agent OS is J&M’s own operating system — which does three things for the business:

1. Proof. Node Zero: we sell the operating-system outcome and we visibly run on one. The demo is the daily driver.
2. The reference build. Agent OS on a VPS behind Cloudflare Access is a repeatable pattern — the credible skeleton for the Company Brain offer (self-hosted second brain per client; see the company-brain-offer). We’ve de-risked the deploy on ourselves first.
3. Capacity. The Concierge base is deliberately uncapped (01 §7a, revised 2026-07-05): the ceiling of a one-person shop is being measured, not asserted, and Agent OS is the instrument that keeps raising it. Every chore it absorbs (drafting, auditing, reporting, filing) is capacity bought back — the killer variable in the unit economics (std-opportunity-analysis: capacity is the constraint, not demand). The §7a tripwire log lives here too: when two tripwires fire in a week, the hiring ladder engages.

Guardrails (what stays human, what stays locked)

• No public RCE. Agent OS has no built-in auth and its API routes shell out to CLIs — Cloudflare Access in front is non-negotiable; the app never binds 0.0.0.0. (Full threat model in the deploy runbook.)
• Human keeps the judgment calls — the same three invariants as client builds (07, Part 2): sending a third party, spending money, and publishing are always human-gated. Agent OS drafts; a person sends.
• Cost discipline: free/cheap tier (OmniRoute / FCC / Ollama) does the volume; paid models are reserved for work where quality pays for itself.

The 90-day plug-in (map to the validation gate)

The master plan’s gate: don’t scale until 3 paying Concierge clients prove the funnel (01, §90-day). Agent OS supports that gate without distracting from it — pick the 3–4 tabs that touch revenue and ignore the rest until the funnel’s proven:

1. Claude control room — run discovery → findings → proposal (Pitch/Close).
2. Hermes Goal Mode — the standing tool-intake + monthly re-audit chore (Find/Ops).
3. SEO + Radar — the inbound/content motion (Find).
4. Goals + Memory Galaxy — the instrumented hunt→…→deploy engine + one place the whole machine reports (Ops).

Everything else (studios, Paperclip, extra-model council) is a buffet, not a checklist — reach for it when a specific job calls for it.

Cross-references

• Prime Directive #3 + positioning — 01-business-plan.md §0.3, §positioning
• The internal engine (hunt→evaluate→categorize→master→deploy) — 00-thesis.md
• The pipeline this fronts (Find→…→Expand, findings-JSON spine) — 07-automation-connector-map.md, Part 1
• Solution stack (client-facing tool choices) — 06-solution-stack/stack-table.md
• Prospecting automation (pipeline.py + Twenty CRM) — prospecting-pipeline memory, ~/projects/ai-business-prospects/
• Deploy runbook + threat model — agent-os-agentic-os memory, personal-ai-os/infra/agent-os/

Naming note: the CRM was EspoCRM until 2026-07-01, now Twenty (self-hosted, on the platform stack). Sections 06 and 07 are reconciled to say Twenty; if any older doc still says EspoCRM, read it as “the CRM (now Twenty).”