Video summary

Hermes Agent Desktop: Full Setup + Real Use Cases

Main summary

Key takeaways

Product Review

Product being reviewed

Hermes Agent Desktop (desktop app) for using Hermes agents—including features like sessions, profiles (separate agents/models), artifacts (second-brain style organization), skills/tools management, cron job scheduling, and sub-agents—with claims of better efficiency than OpenClaw and earlier Telegram/CLI-based workflows.

Key features mentioned

Desktop session management (multi-session / context separation)

  • Each message goes into a new “session” (vs Telegram workflows that require manual/thread setups).
  • Helps avoid “polluted context” that makes token/costs spike.
  • Includes pinning/organization and folders for sessions.

Profiles = multiple agent instances with different models/skills

Profiles are described as separate Hermes agents with their own:

  • skills
  • personality (“soul.md”)
  • memories
  • session/chat history

Desktop UI makes switching profiles easy (instead of terminal commands).

Example profiles mentioned:

  • default
  • Billy
  • coder
  • GPT-me’s
  • librarian
  • Oracle
  • Gwen

Model-to-profile switching for cost/performance

Recommended thinking: choose profiles based on model strengths/weaknesses (not 1 profile per job role).

Examples:

  • Opus 48 for expensive/high-level strategy
  • GPT-me’s (ChatGPT 5.5) for coding with “huge limits”
  • Qwen local model for research and “free” local runs

Artifacts (centralized file/link/image organization)

  • Stores links, images, files in one searchable place.
  • Used like an automated “second brain”: send a link once and it gets organized (with an “organization agent” like a librarian).

Skills & tools management

  • Out-of-the-box: “over 150 skills” that can increase cost by adding context.
  • Desktop UI lets users turn off unnecessary skills to save money.
  • Mentions Hermes may generate skills in the background; interface exposes what’s installed/generated.
  • Tool sets are described as a new concept: groups of tools/skills for complex tasks.

Messaging setup without CLI

  • Desktop UI makes it easier to set up messaging services (previously required CLI).
  • Goal: “never have to go into your terminal again.”

Cron scheduling with visibility/verification

  • Desktop cron section shows scheduled jobs so users can confirm routines actually exist.
  • Includes creating cron manually via UI.
  • Example use: daily “morning brief” delivered to desktop.

Reverse prompting to improve cron prompts

Advice: do a brain dump of interests/goals/skills, then use reverse prompting to generate the best cron-job prompt for the user.

Example “morning brief” prompt output includes:

  • web search for last 24 hours
  • structured sections (AI news, marketing/investing, tech news)
  • formatting guidance (bold, bullets) to avoid unreadable dumps

Context controls

  • Mentions a context window indicator and using “/ new” to start a new session to clear context and save money.

Dynamic model swapping

  • Claims Hermes makes it easier to swap models and “swap in” new models when released (unlike OpenClaw, where models might be “hardcoded” and require waiting).

Sub-agents vs profiles

  • Sub-agents: copies of the main agent (same skills/personality), used for parallel work on one skill set.
  • Profiles: different skill sets/contexts/memories, used when different capabilities are required.

Action/use-case example: automated business opportunity scanning

A cron job runs every 20 minutes using Qwen locally to scan:

  • Reddit
  • X (Twitter)

It produces a “custom dashboard” with:

  • identified “challenges/opportunities”
  • source threads
  • reasoning why the speaker is positioned to solve them
  • optionally a “prototype/prototype button” to build an initial micro-SaaS prototype

Local hardware recommendation (to run local models)

Suggested options:

  • Mac Studio (preferred for UX; more unified memory; but sometimes sold out / limited memory)
  • DGX Spark (preferred “plug-and-play” if Mac Studio isn’t available)

Additional detail:

  • Price update: DGX Spark raised to ~$4,800
  • Used for running local models like Qwen (e.g., Qwen 27B cited) and newer open-source models.

Pros (as stated)

  • Better than Telegram/CLI workflows for using agents (more “Apple-esque” desktop experience).
  • Session management reduces cost by preventing context pollution.
  • Profiles + UI switching make multitasking practical.
  • Artifacts makes it easy to centralize “second brain” materials (links/images/files).
  • Skills/tools toggle UI helps reduce unnecessary cost overhead.
  • Cron UI improves reliability: users can see scheduled jobs and be confident they’ll run.
  • Dynamic model switching is easier than the comparison product.
  • Sub-agents enable parallel execution for multi-feature builds.
  • Local model usage can reduce ongoing LLM spend (claimed “free” when using local compute).

Cons / limitations mentioned (or implied)

  • Potential cost blowups are real if users don’t manage sessions/context correctly (speaker cites complaints like paying very high monthly costs).
  • DGX Spark costs ~$4,800 (presented as expensive, but framed as an investment).
  • Some “idea browsing / AI slop” output may not be pretty; value is framed as opportunity leads/prototyping.
  • Critique of “profile per every job role”: may be inefficient/costly and increases decision overhead (speaker prefers model-based profiles instead).

Comparisons made

Hermes vs OpenClaw

  • Hermes described as the “best AI agent experience” and a step forward from OpenClaw.
  • Claim: Hermes updates are more polished, tested, focused, and less likely to break.
  • OpenClaw described as more “Android route”: more features, less testing, potentially unreliable and breaking updates.
  • Hermes also claimed to be more flexible for model switching.

Telegram/Signal/CLI workflows vs Hermes Desktop

  • Telegram requires complicated thread/group setup for multi-context use; Hermes Desktop makes sessions simpler.
  • Hermes Desktop avoids CLI steps for messaging service setup and cron verification.

“Paperclip-style” role profiles vs model-based preference

  • Speaker prefers profile selection based on model strengths rather than 50 role-specific agents (argues that “AI agent in every role” can add inefficiency/cost).

User experience notes (how it feels / usability)

  • Described as highly organized and visual, positioned as friendlier than CLI-heavy usage.
  • Switching between profiles is fast via UI instead of terminal commands.
  • Cron jobs are more dependable because they’re visible in the app.
  • Provides “second brain” organization (artifacts) without manual filing work.

Unique points mentioned (exhaustive list)

  1. Hermes Desktop launched recently; framed as replacing prior Telegram workflows.
  2. Desktop sessions create context separation automatically; avoids Telegram thread complexity.
  3. Proper session management reduces cost; bad context causes expensive bills.
  4. Profiles are multiple separate agents with their own skills/personality/memories/history.
  5. Desktop UI organizes profiles (quick switching/multitasking).
  6. Suggested approach: choose profiles by model strengths (Opus vs GPT vs Qwen) not by job roles.
  7. Coding: ChatGPT 5.5 (GPT-me’s) claimed better for coding with more limits.
  8. Strategy: Opus 48 for in-depth planning.
  9. Research: Qwen local for “free” scanning.
  10. Artifacts centralize links/images/files; supports searching and organizing.
  11. Artifacts supports “MacGyver second brain” behavior more easily.
  12. Hermes felt more “Apple-esque” than OpenClaw (“pleasant for average person”).
  13. Skills: out-of-box 150+ skills can increase context/cost.
  14. Desktop lets users toggle skills to save money and increase efficiency.
  15. Hermes generates skills in background; UI exposes them.
  16. Tool sets group multiple skills/tools for complex tasks (new concept).
  17. Desktop enables messaging setup via UI rather than CLI.
  18. Cron/jobs reliability improved because UI shows scheduled tasks.
  19. Cron prompting: “reverse prompting” + brain dump to generate better scheduling prompts.
  20. Example cron prompt includes strict “last 24 hours” web search and structured formatted output.
  21. Context window management: can reset context with “/ new” to save money.
  22. Hermes allows easier model switching and dynamic swapping when models update.
  23. Agent window shows multiple sub-agents running.
  24. Sub-agents are copies of main agent; profiles are distinct agents with different contexts/skills.
  25. Example sub-agent use: micro-SaaS with 5–6 features built in parallel.
  26. Example multi-profile use: research → scripting → thumbnail image creation uses different models.
  27. Real use case: solopreneur opportunity scan via cron every 20 minutes.
  28. Opportunity scan uses Reddit + X to find challenges and provides source threads.
  29. Agent output includes rationale why user can solve it and may generate a prototype.
  30. Local models recommended: Mac Studio (UX) vs DGX Spark (plug-and-play).
  31. DGX Spark price cited: $4,800 (after increase).
  32. Hardware ROI framing: treat as investment, compare to subscription costs.
  33. Advice: don’t just “play”—use agents to solve others’ challenges to make money.

Speakers / differing views (at end)

  • Greg Eisenberg (host)

    • Pushes for explanations of why to get Hermes Desktop, how to make money and replace OpenClaw/Telegram usage.
    • Asks about skills/tools, cron wording, and overall unanswered questions.

  • Alex Finn (main explainer)

    • Provides the bulk of product walkthrough and optimization tips:
      • session management for cost control
      • profiles and model-based workflow
      • artifacts as second-brain centralization
      • skills toggling/tool sets
      • cron reliability + reverse prompting
      • sub-agents vs profiles
      • real solopreneur opportunity scanning example
      • local hardware (DGX Spark/Mac Studio) recommendations and pricing/ROI framing
      • Hermes vs OpenClaw reliability/polish comparison

Concise verdict / recommendation

Recommended for people who want a more organized, cost-controlled, desktop-first workflow for Hermes agents—especially if you’re currently using Telegram/Signal/CLI or OpenClaw.

The strongest value claims are session/profile management for efficiency, artifacts for organization, and UI-based cron reliability.

If you plan to run local models, be prepared for hardware costs (~$4,800 DGX Spark cited), but the video argues it can pay off via reduced ongoing LLM spend and increased opportunity-generation/prototyping.

Original video