Agent workflows
Configured model access, tool calling, scheduled tasks, reminders, file delivery, and service actions across several personal use cases.
AI Agents
Hermes Agent
A personal AI agent that connects chat, tools, scheduled workflows, and structured data on a privately managed home server.
- Client
- Personal project
- Role
- AI Agent
- Year
- 2026
- Status
- Active development
Hermes Agent dashboard showing the Debian host and connected service configuration
Screenshot 1 of 3
Overview
One agent across the systems I use every day
Hermes Agent is a personal AI system that I host and operate on a repurposed laptop running Debian 13. I use it for career planning, cash flow records, learning, research, and recurring automation through Telegram or a web dashboard.
The agent can call tools, work with external services, deliver files, and run scheduled workflows. Google Sheets, Google Calendar, Notion, and GitHub still hold the underlying records. Hermes provides a conversational and operational layer across those systems instead of replacing them with another database.
I configured the private server, remote access, service integrations, and workflow rules. I also created dedicated skills for personal cash flow and Telegram file delivery, with verification after actions that change external data.
Technology stack
Agent and models
- Hermes Agent
- GPT 5.6 Sol & Luna
- Agent Skills
- MCP
- Tool Calling
Interfaces and services
- Telegram
- Web Dashboard
- Google Workspace
- Notion
- GitHub
- Browser Automation
- Speech to Text
Infrastructure and access
- Debian 13
- Headless Home Server
- Tailscale
- SSH
- Ed25519
ProblemPersonal workflows were spread across separate tools
Job applications, financial records, learning schedules, notes, files, and reminders live in different services. Each service is useful on its own, but completing a workflow often requires moving between several interfaces and repeating the same context.
I wanted one assistant that could work across those services without turning them into loosely synchronized copies. The system also needed private remote access and a safer way to handle actions that change records.
Keep the existing sources of truth
Applications and finance records stay in Google Sheets, schedules stay in Google Calendar, and learning notes stay in Notion.
Support actions beyond conversation
Hermes can call tools, run scheduled work, send files, and update connected services instead of stopping at a generated response.
Preserve control over important choices
Workflows such as learning briefings retain explicit actions to start, postpone, or skip the next step.
ContributionBuilt and operate the personal agent stack
I set up the Hermes runtime on a headless Debian 13 server, connected its Telegram and browser interfaces, and configured private administration through Tailscale and SSH.
I connected the agent to the services used by each workflow and defined how it should read, update, and verify structured records. I also created custom skills for transaction handling and Telegram file delivery.
Custom skills
Created dedicated skills for personal cash flow and Telegram file delivery rather than relying on one general prompt for every task.
Private infrastructure
Repurposed a laptop as a Debian server without a desktop environment, then secured remote administration through a private network and SSH keys.
ArchitectureA private agent connected to the systems that hold the data
Requests enter through Telegram or the web dashboard. Hermes selects a configured model, skill, or tool, then works with the relevant external service. GPT-5.6 Sol and Luna provide the reasoning and language layer for different task configurations.
The runtime operates on a headless Debian 13 home server. Tailscale provides a private network path, while SSH with Ed25519 keys supports remote administration. Actions that modify external data follow an execute, verify, and confirm sequence before Hermes reports success.
Interfaces
Telegram
Chat, transaction entry, reminders, briefings, workflow triggers, and file delivery.
Web dashboard
A browser interface for conversations, configuration, and richer agent controls.
Agent layer
Hermes runtime
Routes each request through the required model, skill, and tool sequence.
Models
GPT-5.6 Sol and Luna provide reasoning and language generation through Codex OAuth.
Skills and tools
Package workflow instructions, tool calls, scheduled tasks, and verification rules.
Connected systems
Google Workspace
Stores job applications, financial records, learning schedules, and related documents.
Notion and GitHub
Hold learning notes, curriculum material, and project information.
Browser and speech tools
Extend workflows that require web interaction or spoken input.
Private infrastructure
Debian 13 home server
Runs Hermes without a desktop environment on a repurposed laptop.
Tailscale
Connects trusted devices to the server over a private network.
SSH
Supports remote administration with Ed25519 keys protected by a passphrase.
ResultsHow I use Hermes today
Hermes is already part of my career planning, personal finance, learning, research, and recurring automation. I use Telegram for routine actions and the web dashboard when I need more room for interaction or configuration.
I have not run a formal evaluation of task success, latency, or operating cost yet. This section therefore describes the workflows I currently use rather than presenting performance claims.
Career workflows
Researches opportunities, evaluates role fit, prepares application material, manages CV variants, and works with the application tracker.
Cash flow workflows
Turns conversational transaction input into structured records with validation, duplicate checks, transaction IDs, audit logs, and verification after each write.
Learning and recurring work
Coordinates curriculum, schedules, notes, reminders, review sessions, spaced repetition, briefings, and selected browser tasks.
LearningsPrinciples shaping the system
Keep records in the system that owns them
The agent coordinates Google Sheets, Calendar, Notion, and GitHub without treating every store as a synchronized copy.
Verify writes before reporting success
A convincing model response does not prove that an external record changed. The workflow reads the result back before confirming the operation.
Retain human choices in recurring workflows
Automation handles reminders and preparation, while actions such as starting, postponing, or skipping a learning session stay explicit.
Next project
ICCN Chatbot
An ICCN knowledge assistant with a gamified flow for collecting creative-economy participant data.