Skip to main content

HERMES AGENT

Self-hosted Hermes Agent by Nous Research (MIT) set up with a Coordinator and 4 subagent roles (EA, CoS, OM/BizOps, PA), Policy Interceptor gates, and 3-tier data governance.
Executive Assistants
Active Personal Ops
CATEGORY
Executive Assistants
STACK
Nous Research (MIT) · Coordinator & MCP
STATUS
Active Personal Ops
Topology
Single Thread
Write Policy
Manual Checks
Low-Impact Ops
Manual Repetitive
Uncertainty
Silent Guessing
Data Governance
Flat Context
Error Handling
Cascading Fails
Memory & Skills
Stateless Chats
Tool Surface
Ad Hoc Wrappers
Topology
Coord + 4 Roles
Write Policy
Policy Interceptor
Low-Impact Ops
Pre-Approved Cat.
Uncertainty
Proposal Halt
Data Governance
3-Tier Partition
Error Handling
Failure Isolation
Memory & Skills
Tiered + Skills
Tool Surface
10 Categories / MCP
MEDIA FORMAT NOTE

Presentation film: 15-second motion overview showing the self-hosted Hermes workflow and boundary posture (not a capture of private corporate systems).

Attribution: Open-source software by Nous Research (MIT License); Coordinator topology, governance, and operation by Joaquin Villarreal.

Overview & Problem Addressed

Hermes Agent is an open-source AI agent framework by Nous Research (MIT License) that includes Hermes-2-Pro and Hermes-3 foundations, SQLite/FTS5 session storage, Markdown-curated memory, reusable skills, CLI and messaging gateways, and MCP tool routing. This project documents how I deploy and configure Hermes in a self-hosted personal environment around a Coordinator Mode blueprint with four specialized subagent roles modeled on executive support: Executive Assistant (EA), Chief of Staff (CoS), Office Manager / BizOps (OM), and Personal Assistant (PA).

Capabilities & Workflow Execution

Every tool call passes through a Policy Interceptor Gate with broad read access on Tier 1 operational sources and writes blocked by default. Write actions proceed only under two deterministic conditions: either the call matches an entry in a Pre-Approved Automatic-Action Catalog for reversible, low-impact prep (drafting unsent replies, placing tentative calendar holds, triaging tasks, or logging receipts below threshold under an EA audit log), or it receives explicit human approval (EA sign-off for Tier 1/3; CEO sign-off for Tier 2). Ambiguity triggers the Proposal halt rule for human confirmation, and any subagent timeout, error, or mismatch is quarantined under the Failure Isolation Rule without synthetic guessing.

Executive Assistant Support & Application

A three-tier data model governs daily operations. Tier 1 covers scheduling, correspondence prep, and travel logistics. Tier 2 covers finance, treasury, and payments with restricted reads and zero automatic payouts (100% CEO approval after EA filtering). Tier 3 restricts HR, legal, Board packs, and confidential files, keeping raw contents out of shared vector stores and escalating by metadata ticket only. Alongside Minto morning briefs and calendar protection, CoS tracking maintains a delegated-commitment ledger ('Monkey Tracking') while hiring, termination, legal, and contractual choices stay human.

Configured Implementation

The integration blueprint covers 10 connectable tool categories: Calendar, Communication, Finance, CRM, Docs/Knowledge, BI/Analytics, Automation/Orchestration, Research/Web Intelligence, Security/IAM, and Personal Logistics. It separates verified Official Core APIs from isolated Community Integrations and enforces Startup Re-Verification against live runtime manifests. Scenario blueprints act as reference workflows only; personal self-hosted instances stay isolated from corporate credentials and private company stores.

Operational State & Availability

Actively operated in my self-hosted personal environment, with the multi-role executive setup maintained as sanitized documentation. Nous Research (MIT License) holds attribution for the base software; Joaquin Villarreal handles configuration, permission design, and executive workflow integration. The 15-second video is a presentation clip rather than corporate footage, and private configs, keys, and internal prompts stay offline.

AVAILABILITY & DIRECT CONTACT

Want to discuss this project or set up a similar workflow? Write to jevillarreal170@gmail.com and I will gladly share what can be shown.

Public repository links stay unpublished while private configs, credentials, and environment records remain offline. No confidential company data or invented dates appear here.