An autonomous AI agent that browses the web, edits files, runs Claude Code, and works for you 24/7 — all from Telegram.
It's like a remote intern with superpowers — give it a task, it figures out how to do it.
Hermes is a remote intern with superpowers. You give it a task, it figures out how to do it — browsing websites, editing files, writing code, managing WordPress — and reports back. You don't have to hold its hand through every step.
Three superpowers that cover 90% of SEO operations.
Opens Chrome, logs into WordPress, edits pages, takes screenshots, fills forms — hands-on, like a human would.
Reads, writes, searches, and edits files on disk. Manages content, scripts, and project files like Claude Code.
Runs Python, shell scripts, git. Fires up Claude Code CLI for complex coding and reasoning tasks.
Fast and cheap as the primary model — Claude Code gets called when things get hard.
| Trait | Why it matters |
|---|---|
| ⚡ Fast | Near-instant responses, feels responsive |
| 💰 Cheap | Most cost-efficient — keeps API costs low |
| 🎯 Efficient | Handles 90%+ of daily tasks without slowdown |
DeepSeek is the dispatcher — fast, cheap, good at routing. Claude Code is the specialist — slower but smarter, called in when needed.
Why we run Hermes as our primary agent.
| Factor | Hermes | Claude Desktop |
|---|---|---|
| Cron Jobs | Unlimited — recurring tasks run autonomously | None — one-shot only |
| Cross-Session Memory | Retains — searches past sessions | Blank slate every session |
| Self-Improvement | Patches skills, saves memory automatically | Static — no auto-improvement |
| Mobile & Team | Telegram — from anywhere, multi-user | Local — mobile exists but sessions don't persist |
| Multi-Model | Swap mid-task via /model |
Tied to one model per session |
| Tool Ecosystem | MCP + Skills + sub-agents | CLI only — no MCP |
Three slash commands to control the agent.
/model <name>Swap the AI model mid-conversation. Example: /model anthropic/claude-sonnet-4
/restartHard reset across all sessions — Telegram, CLI, cron. Use when stuck or unresponsive.
/skillsList all available skills — reusable workflows the agent can load on demand.
Instruction manuals the agent reads before starting a job — steps, pitfalls, and verification.
wp-onpage-optimizerWordPress on-page SEO workflow
domain-seo-auditFull SEO audit from scratch
sister-site-auditSister site analysis
content-gap-expanderKeyword gap analysis
pbn-link-buildingPBN link building campaigns
project-trackerUpdates Obsidian vault automatically
Skills are never final — tweak as you go. Nothing is final until it ranks.
Each service runs an MCP server. Hermes discovers and uses them automatically — no per-service API keys needed.
Site Explorer, Keywords, Rank Tracker, Backlinks
Search Console — clicks, impressions, CTR, position
SERP, backlinks, keywords, AI optimization
Project tracking CRUD
These work regardless of which skills or MCP servers are loaded.
Navigate, click, type, screenshot, run JS
Read, write, search, patch files
Shell commands, scripts, bg processes
Real-time information retrieval
Analyze images and screenshots
Persist facts across sessions
Recall past conversations
Spawn sub-agents for parallel work
Schedule recurring tasks
Telegram, Discord, email delivery
Create, patch, delete skills
Chain tool calls in Python
Audio output
A local markdown wiki that powers long-term campaigns — searches, trackers, decisions, all in plain text.
Status, deliverables, URLs, checklists — one note per campaign
M4 reviews, audit results, stored as searchable notes
Why we chose certain approaches, what failed, what worked
| Pro | Con |
|---|---|
| Very fast — local markdown, no API calls | Local — on Ric's PC, not cloud-accessible directly |
| Single source of truth — auto-updated by project-tracker skill | ClickUp MCP sometimes down |
| GitHub sync — view-only for team |
How we work with AI — and how we don't.
It executes faster and more consistently than a human — but strategy, judgment, and final decisions are yours.
Better prompts = better results. If it fails, try a different method or model. AI hallucinates — always double-check the output.
Publish → Monitor → Adjust → Repeat. Skills evolve as we learn. If it doesn't move rankings, change it.
Worried about touching a team skill? Ask Hermes to create a personal one — zero risk to shared workflows.
Just say: "Create a skill called michelle-seo-onpage that does X, Y, Z"
Tweak freely — no risk to team workflows
"Use michelle-seo-onpage for this"
Patch, add pitfalls, remove stale steps
Share with team once it proves itself
Everything lives in Ric's office PC under ~/Documents/Obsidian Vault/ — plain markdown, no cloud lock-in.
| Folder / Note | Purpose |
|---|---|
Hermes Usage/ | This doc, onboarding deck, account usage logs |
What is Hermes Agent.md | This document — team onboarding & reference |
account_usage_log.md | API consumption tracking |
Marketing Intelligence/ | Daily AI marketing briefs & competitor intel |
AI Marketing Daily/ | Daily marketing roundups by date |
Sister Sites/ | Per-site project trackers |
JDL688/ | JDL688 sister site trackers & reviews |
PBN Link Building/ | PBN campaign notes & performance reviews |
Aged Domains/ | Aged domain project files |
🔍 Hermes can search any note on demand — just ask.