Frank: My Hermes Setup

Frank is the name of my Hermes setup. The name is funny. The architecture is serious.

Most AI demos focus on the model. Frank is the system around the model: persistent context, reusable skills, scheduled research, model routing, state, verification, and a human approval boundary. The model can change. The operating layer is what makes the work repeatable.

This is a personal lab system. It is not a district-managed AI platform, and the public map intentionally leaves out credentials, private endpoints, internal source names, and organization data.

What flows into Frank

The Bee wearable provides ambient context. The wearable captures conversations and voice notes. A scheduled local pull imports the useful summaries into a private cache, where later workflows can use them as context. The device shown in the map is from Bee's public product page.

Research collectors run quietly. Daily and weekly jobs monitor K-12 AI and edtech, security and compliance, policy, model releases, speaking opportunities, public conversations, and other named sources. Most of those jobs write local evidence artifacts. They do not all notify me.

I work through several surfaces. WebUI and Desktop are useful for longer work. Telegram is the mobile surface. Local tools let Frank read files, run bounded commands, operate approved applications, and verify what changed.

The research system is a filter

The 48 scheduled tasks are not 48 daily interruptions. They are a pipeline:

1. Discover from named source categories.
2. Verify dates, links, provenance, and whether an item is actually new.
3. Write state so the system does not rediscover yesterday and call it fresh.
4. Synthesize only when several sources support a useful pattern.
5. Deliver selectively through the Daily Field Report, research ledgers, speaking packets, knowledge updates, and draft public work.

Silence is an output. If nothing crosses the bar, a dependable system should leave me alone.

Profiles route work to the right model

Frank is not one giant personality wearing six hats badly. Hermes profiles give recurring roles their own instructions, memory posture, tools, and model route.

That routing is deliberate. Model choice follows the work. It is not a loyalty program.

Skills are the reusable judgment layer

A skill is a documented procedure that tells Frank how to do a recurring kind of work. Skills contain triggers, ordered steps, safeguards, verification, and known failure modes. They cover research, documents, publishing, Apple tools, coding, media, agent maintenance, security, and more.

There are 127 enabled skills in this snapshot. The catalog below lists the public-safe ones. Person-specific, internal, and security-sensitive procedures are excluded, because publishing every instruction your agent knows is how a useful diagram becomes an incident report.

Memory, state, and proof

Memory stores durable facts. Skills store procedures. Scheduled tasks execute procedures. State records what already happened. Tools touch the real world.

Verification is the final layer. Frank reads results back, checks artifacts, separates task success from delivery success, and keeps external actions such as publishing, messaging, purchasing, permission changes, and destructive operations behind human approval.

I do not approve an agent because the demo is impressive. I approve a bounded workflow with named inputs, permissions, verification, and an owner.

Build one useful loop

The practical starting point is not 48 scheduled tasks. It is one recurring decision your team already makes.

Define the trigger, inputs, context, decision rule, permitted action, verifier, state, notification threshold, prohibited actions, and escalation conditions. Run it. Inspect the proof. Fix the weak edges. Then expand.

The PowerPoint above walks through that design in 15 slides. The PDF is there because projectors occasionally become agents of chaos.