Verdict: Most people install Hermes Agent, run hermes setup, chat with it in the terminal, and stop there. That's like buying a Swiss army knife and only using the blade. The real power shows up when you configure profiles for isolation, schedule cron jobs for unattended work, register free 295-billion-parameter models through Nous Portal, and let the agent write its own skills from experience. These seven configuration tricks — each verified against the official Hermes Agent docs and the Nous Portal — will take you from casual user to someone running a genuine AI operation.
Last verified: 2026-08-01 TL;DR:
- Run
hermes sessions optimize-storageto shrink your session database by up to 78% (auto-triggers if under 1 GB)- Create isolated profiles with
hermes profile createso different projects don't contaminate each other's memory- Register free models like Tencent Hy3 (295B MoE, 256K context) through Nous Portal at $0 for evaluation
- Schedule unattended jobs with
hermes cron create— they run in fresh sessions with their own skills- Delegate parallel work with
delegate_taskso three sub-agents research simultaneously instead of sequentially- Let Hermes auto-generate skills after a 5+ step task by saying "save what you just did as a skill"
- Paste screenshots with Ctrl+V and the agent analyzes them with vision — no file-saving needed
1. How Do You Shrink the Session Database When It Starts Slowing Down?
Hermes Agent stores every conversation — CLI, Telegram, Discord, Slack — in a SQLite database at ~/.hermes/state.db with FTS5 full-text search. Over months of heavy use, this database balloons, slowing session loads and history search. A storage optimization update introduced a command that compresses the database by up to 78% (about 60% on average) by pruning redundant message metadata and compacting old session records.
Run the optimization manually:
hermes sessions optimize-storage
If your database is under 1 GB, this one-time optimization triggers automatically after updating Hermes — no command needed. For larger databases, run the command once after updating. The process is safe: it preserves all session content and message history; it only reclaims space from internal overhead.
You can check database size before and after with:
hermes sessions stats
This prints total sessions, message counts, per-platform breakdowns, and current database size — so you can measure the savings yourself.
2. What Are Profiles and Why Does Every Pro User Set Them Up?
A Hermes profile is an isolated copy of the agent with its own model, memory, skills, gateway connections, and configuration. Without profiles, every project shares one brain — your coding agent's memory leaks into your marketing agent's context, and gateway tokens collide.
Create a profile for each workflow:
hermes profile create research-bot --description "Web research agent with arxiv and web-search skills"
hermes profile create coder --description "Coding agent with github-pr-workflow skill"
Switch between them with the -p flag:
hermes -p research-bot
hermes -p coder
Each profile lives at ~/.hermes/profiles/<name>/ and has its own config.yaml, MEMORY.md, SOUL.md, skills/ directory, and cron/ directory. This means your research agent can have a different model, different system prompt personality, different scheduled jobs, and different memory from your coding agent — all on the same machine, no extra installs. This isolation pattern is what separates disciplined agent setups from the session sprawl that quietly halves your output — each profile keeps context focused and prevents bleed.
Common pitfall: If two profiles use the same Telegram bot token, one will stop receiving messages. Each profile needs its own gateway configuration. Run hermes gateway status per profile to verify.
3. How Do You Run Hermes Agent on a Free 295B Model?
Hermes Agent supports almost any LLM provider — OpenAI, Anthropic, Google, local models via Ollama, Hugging Face, and Nous Portal, which launched April 27, 2026 and bundles access to 300+ models behind a single subscription.
In July 2026, Nous Research announced that Tencent's Hy3 — a 295-billion-parameter Mixture-of-Experts model with a 256K context window — would be free on the Nous Portal. Hy3 was built specifically for agentic use: coding, reliable tool-calling, and reasoning. It achieved GPQA Diamond 90.4 and SWE-bench Verified 78.0, placing it in the frontier tier.
To wire Hy3 into Hermes:
# 1. Authenticate with Nous Portal (browser opens once)
hermes auth add nous
# 2. Set the model — note the :free suffix
hermes config set model tencent/hy3:free
hermes config set provider nous
The :free suffix is critical: the plain slug tencent/hy3 hits the paid meter. tencent/hy3:free routes through the free tier at $0. This is the same free-routing pattern that works for connecting Hermes to the Buzz Nostr workspace — one login, no API key copying.
Which free model should you pick? Testing by the community and benchmark data show:
| Model | Parameters | Context | Best For | Cost |
|---|---|---|---|---|
| Tencent Hy3 | 295B MoE (21B active) | 256K | Coding, tool-calling, agent workflows | Free on Nous Portal |
| Laguna S 2.1 | 118B (8B active) | 262K | Hard reasoning tasks | Free via OpenRouter / Nous Portal |
| Nous Hermes 3 405B | 405B | 131K | Roleplay, general agentic tasks | Free tier (deprecating) |
Hy3 is the standout for agent work — reliable tool calls, clean code generation, and the largest free context window. Laguna S 2.1 edges it out on pure reasoning benchmarks but is less consistent on multi-step agent tasks. Neither matches top paid models like Claude Opus 4.6 or Kimi K3 on the hardest tasks, but for $0 in a cloud deployment, they're more than sufficient for most workflows.
Where to verify: The Nous Portal info page lists current free models. Free tiers rotate, so check before relying on a specific model for production.
4. How Do You Schedule Jobs That Run While You Sleep?
Hermes has a built-in cron scheduler that runs agent tasks on a schedule — daily briefings, nightly backups, weekly audits, or monitoring scripts — with no human present. Cron jobs run in fresh agent sessions (no memory of your current chat), so prompts must be fully self-contained.
Create a job via natural language in a chat session:
Set up a cron job that runs every morning at 9am, checks my GitHub notifications, summarizes the important ones, and sends the summary to me on Telegram.
Or from the CLI:
hermes cron create "0 9 * * *" \
"Check GitHub notifications for shamuddin. Summarize anything marked urgent or involving security. Ignore automated dependency bump bots." \
--name "Daily GitHub digest" \
--deliver telegram
Schedule formats:
| Format | Example | Meaning |
|---|---|---|
| Cron expression | 0 9 * * 1 |
Every Monday 9 AM |
| Interval | every 2h |
Every 2 hours |
| One-shot ISO | 30m |
30 minutes from now |
Script-only mode runs bash or Python without invoking an LLM at all — zero token cost per tick. This is ideal for monitoring:
hermes cron create "0 * * * *" \
--no-agent \
--script ~/.hermes/scripts/disk-monitor.sh \
--name "Disk space watchdog" \
--deliver telegram
The script's stdout is delivered verbatim. Empty stdout means silent — nothing is sent. A non-zero exit code sends an error alert. This is the cheapest way to run a 24/7 monitoring operation without burning API credits.
Key constraint: Cron-run sessions cannot recursively create other cron jobs — this prevents runaway scheduling loops. If a cron job needs to trigger additional work, it should create a Kanban task or send a message that a human (or another scheduled job) picks up.
5. How Do You Delegate Work to Multiple Agents at Once?
Hermes supports parallel sub-agent execution through delegate_task. Instead of asking one agent to research three topics sequentially, you spawn three sub-agents — each with its own conversation, terminal session, and toolset — and they work simultaneously. Only the final summary returns to your context.
Typical pattern:
I need research on three topics in parallel:
1. Compare vLLM vs SGLang vs TensorRT-LLM for serving LLMs
2. Find the latest benchmarks for Qwen3 Coder on SWE-bench
3. Identify the cheapest API providers for Claude Opus 4.6 in August 2026
Hermes' agent loop automatically:
- Spawns isolated sub-agent sessions
- Injects relevant context and skills into each
- Runs them concurrently
- Collects results and returns a consolidated summary
When to delegate:
- Research-heavy tasks that would flood your context with raw web page data
- Code review + testing + documentation in parallel
- Multi-file edits where each file is independent
When not to delegate:
- Mechanical multi-step work with no reasoning — use
execute_codeinstead - Tasks needing user interaction — sub-agents cannot ask questions mid-run
- Single tool calls — just call the tool directly
6. How Do You Make Hermes Write Its Own Skills?
Skills are Hermes' procedural memory — reusable, multi-step workflows packaged as markdown files with optional scripts. The official tips guide says: if a task takes 5+ steps and you'll do it again, ask the agent to create a skill.
After completing a complex task successfully, say:
Save what you just did as a skill called deploy-staging.
Hermes writes a SKILL.md file to ~/.hermes/skills/deploy-staging/ with the full procedure — trigger conditions, numbered steps, exact commands, pitfalls, and verification steps. Next time, invoke it with:
/deploy-staging
The agent loads the full procedure into its system prompt and follows it. Skills can also include supporting files: scripts (scripts/), reference docs (references/), and templates (templates/).
Skill management commands:
/skills— browse available skills/deploy-staging— invoke a skill by name- Ask the agent to patch a skill if you find a better approach: "Update the deploy-staging skill — we switched from npm to pnpm"
The most powerful pattern: let skills compound. Each time you do something hard, save it. Within a month, you have a library of 20-30 skills tailored to your exact workflow. The agent gets noticeably better because it's not rediscovering workflows from scratch.
7. How Do You Use Hermes Vision Without Saving Files?
Hermes supports clipboard image paste in the CLI. Instead of saving a screenshot to a file, loading it, and referencing the path, you press Ctrl+V and the agent analyzes the image directly with its vision capabilities.
Use cases:
- Paste an error screenshot — the agent reads the stack trace and debugs
- Paste a UI mockup — the agent generates the HTML/CSS
- Paste a chart — the agent extracts the data points
- Paste a diagram — the agent explains the architecture
This works in the CLI on Linux, macOS, and WSL2. The agent's active model must have vision support (most frontier models like Claude, GPT-5, and Gemini do; Hy3 and Laguna are text-only).
Other CLI power-user shortcuts:
| Action | Shortcut |
|---|---|
| Multi-line input | Alt+Enter or Ctrl+J |
| Interrupt mid-response | Ctrl+C once |
| Force exit | Ctrl+C twice within 2 seconds |
| Slash command autocomplete | Type / + Tab |
| Resume last session | hermes -c |
| Resume by title | hermes -c "my project" |
| Cycle verbose modes | /verbose (off → new → all → verbose) |
How Do You Stack These Tricks Into a Full System?
The real power emerges when you chain these configurations together. Here's a practical 24/7 operation:
- Profile your research agent (
hermes -p research) with Hy3 as the model (free, 256K context) - Schedule a daily 9 AM cron job that runs a research sweep and delivers results to Telegram
- Delegate the research into three parallel sub-agents (trends, competitors, audience)
- Auto-skill the best findings into a reusable
/weekly-researchskill - Memory stores durable facts about your industry so the agent starts each session already informed
- Session optimization keeps the database lean after months of unattended jobs
The difference between a terminal chatbot and an operation is configuration. Each trick above takes under five minutes to set up. Together, they turn Hermes into a self-improving system that compounds value over time — the same architecture behind a voice-activated agent OS built on free models.
What This Means for You
If you're using AI for your work, a small business, or building a product, Hermes Agent's configuration layer is where the leverage is. A $5/month VPS running Hermes with a free model, three cron jobs, and a handful of skills can replace a surprising amount of human busywork — news monitoring, research synthesis, code review, report generation. The trick is to start small: set up one cron job this week, one profile next week, and let skills accumulate naturally from your real work. Don't try to configure everything at once; configure one thing, verify it works, and add the next.
For more on Hermes Agent capabilities — including how to connect it to messaging platforms, run local LLMs, or set up wake-word voice activation — the official docs are the authoritative reference. If you want a completely free setup with no Portal account, the Omniroute + OpenCode guide walks through an alternative path.
FAQ
Q: What is Hermes Agent? A: Hermes Agent is an open-source (MIT) autonomous AI agent from Nous Research. It runs on your own server (Linux, macOS, WSL2, Termux), supports 75+ skills, 27+ tool categories, persistent memory, and connects to Telegram, Discord, Slack, WhatsApp, Signal, and email. The current release is v0.5.0 (March 2026).
Q: How much does Hermes Agent cost? A: The software is free and open source. You pay only for the model API you use. Free models like Tencent Hy3 (295B) are available on Nous Portal at $0 for evaluation. A $5/month VPS can run Hermes with a free model indefinitely.
Q: How do I install Hermes Agent in 2026?
A: On macOS, Linux, WSL2, or Termux, run curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash then run hermes setup. On Windows, use the PowerShell installer or desktop app. The installer handles Python, Node.js, ripgrep, and ffmpeg automatically.
Q: What is the difference between Hermes Agent and a regular chatbot? A: A chatbot answers questions. Hermes Agent has persistent memory, writes its own skills from experience, runs scheduled cron jobs, delegates parallel work to sub-agents, connects to messaging platforms, and remembers every conversation in a searchable SQLite database. It improves over time rather than resetting each session.
Q: Can Hermes Agent run on a free model?
A: Yes. Configure it with tencent/hy3:free through Nous Portal, or use any free OpenRouter model. Free models are sufficient for most agent work — coding, research, content generation — though they trail top paid models on the hardest reasoning tasks. Check the Nous Portal dashboard for current free model availability.
Q: How do I stop my session database from growing forever?
A: Run hermes sessions optimize-storage to compress it by up to 78%. Databases under 1 GB auto-optimize after updating. You can also run hermes sessions archive --older-than "30d" to archive old conversations, and hermes sessions prune --older-than "90d" to delete expired sessions entirely.
Q: Can multiple Hermes profiles use the same messaging platform?
A: No — each profile needs its own bot token for each platform (Telegram, Discord, etc.). If two profiles share a token, one will stop receiving messages. Run hermes gateway status per profile to verify each connection is healthy.
Discussion
0 comments