UpCounting Mission Control — Architecture Map

Layer 01 Knowledge — the brain The what and why · human-readable · version-controlled

Notion

Source of Truth

Single canonical home for everything humans (and agents) need to read. SOPs, brand voice, strategy docs, brain pages, and — critically — every skill prompt stored in a versioned page template so any agent can pick up and run it.

SOPs & brand — how we work, how we sound, what we promise
Strategy & brain pages — narrative context, hiring, positioning, philosophy
Skill prompts — versioned page template per skill: description, when to use, inputs, prompt body, outputs, owner tool, changelog
Decision context — long-form rationale that backs structured decisions in Supabase

read · Hermes read · Perplexity read · Claude Code write · Perplexity (diff → approve) write · Abir cache · 5min in Hermes

Layer 02 Data — the nervous system Canonical operational truth · real-time · queryable by everything

Supabase

Canonical DB

Postgres + real-time + edge functions. Receives a duplicate of source data from connected tools so we can track what Pipedrive (and others) don't — stage-change timestamps, event history, attribution — and so the dashboard can query one place instead of fanning out.

Skill registry — pointer to the Notion page, owner tool, current version, last run, metrics
Run events — every LLM execution logged richly: provider, tokens, cost, duration, savings claim
Decision log — structured, supersedes-chained, queryable; the bridge between thinking and execution
Mirrored source data — Pipedrive deals, ClickUp tasks, Airtable rows, with full event history
Connector status, alerts, costs, subscriptions, memory index

read · Mission Control (real-time) read · Hermes read · Perplexity write · Hermes (run events, decisions) write · Edge Fns (ingestion) write · Perplexity (decisions)

Layer 03 Execution — the doers Divided by nature of work · LLM vs. deterministic

Hermes

Chief of Staff

The runtime for everything LLM-driven. Owns memory, profiles, and long-lived agents. The only voice that talks to humans on Slack.

Reads skill prompts from Notion (cached)
Reads/writes state in Supabase
Runs agents by profile: Inbox, CRM, Content, Ops…
Handles human-in-the-loop via Slack
Reports up; flags exceptions

in · Notion in · Supabase in · Slack out · Slack out · Supabase

Claude Code

Plumber

Builds deterministic plumbing: ingestion pipelines, integrations, Mission Control itself, schema migrations. Nothing LLM-at-runtime lives here.

Supabase edge functions (webhook listeners)
pg_cron pollers for non-webhook sources
Custom integrations, scrapers, scripts
Mission Control Next.js app
Triggers Hermes agents when events need a brain

in · Notion (specs) out · GitHub out · Mac mini out · Supabase schema

⌒

Schedulers

Heartbeat

The clockwork that fires recurring work. Split between Supabase's pg_cron (for data jobs) and Hermes's own scheduler (for agent runs).

pg_cron — ingestion pollers, nightly rollups, backups
Hermes cron — agent runs, digests, periodic reviews
All schedules visible in Mission Control

deterministic observable

Layer 04 Interface — where you meet the system Push via Slack · Pull via Mission Control

Slack

Inbox

Bidirectional with Hermes only. Work comes to you here; you respond inline; Hermes proceeds. Mission Control never talks to Slack directly — it routes through Hermes so the voice stays consistent.

Inbound — mentions, DMs, slash commands → Hermes interprets and acts
Outbound — human-in-the-loop asks, alerts, digests, status pings from Hermes
Channels: #mc-alerts, #mc-queue, DMs for critical
Single aggregated queue across every agent — one inbox, not five

in · Abir + team in · Hermes out · Hermes

Mission Control

Dashboard

Pull-only. You go here deliberately to look at trends, scorecards, agent activity, spend, ROI, and the knowledge graph. It does nothing on its own — it shows.

Next.js 15 · reads Supabase real-time
Mobile-responsive · hosted on Mac mini · Cloudflare Tunnel
Dashboards, scorecards, decision browser, ⌘K search
Demo mode for screenshots / walkthroughs

in · Supabase (real-time) read-only mac mini · tunnel

Layer 05 Thinking — the partner Reasoning · drafting · personal one-offs · writes back to Notion + Supabase

Perplexity Computer

Thinking Partner

Where Abir reasons, plans, drafts, and handles personal one-offs. Outputs flow back into the operational stack so thinking doesn't get stranded.

Strategy & journaling — long-form thinking with full connector + memory access
Content work — landing pages, scripts, performance analysis
Decisions → written as structured rows in Supabase decision log
Narrative context → proposes diffs to Notion brain pages; Abir approves in Slack
One-off agentic tasks — research, drafts, artifact generation

in · all connectors in · Notion (read) in · Supabase (read) out · Supabase decisions out · Notion (via approval)

Information flow

Data Knowledge Human comms

Routing rules

If

Work needs a model at runtime → Hermes
If

Work is deterministic code → Claude Code
If

Work is a personal one-off or strategic thinking → Perplexity
If

You need to look at the system → Mission Control
If

The system needs to reach you → Slack via Hermes
If

A decision needs to stick → Supabase decision log (+ Notion mirror)
If

A piece of knowledge needs to be reusable → Notion

Hard invariants

Never

Mission Control does anything. It only shows.
Never

Mission Control talks to Slack directly. Hermes is the only voice.
Never

Run an LLM skill outside Hermes in production (except Perplexity for your own one-offs).
Never

Duplicate skill prompts. Notion is the single source; agents cache.
Always

Log every run event richly to Supabase — attribute later, never reconstruct.
Always

Notion edits from Perplexity go through diff → approve → write.
Always

Decisions supersede, never delete. The chain is the audit trail.