Should You Self-Host Hermes or Use Agent37? A Decision Framework for Claude Code Teams

Once you have decided you want Hermes Agent running somewhere, you have to decide where. Nous Research offers Agent37, a hosted Hermes service starting at $3.99/mo. You can also self-host on a $4-8/mo VPS, Modal, Daytona, or your own Kubernetes cluster. The sticker prices look comparable. The real decision axes are data locality, compliance, update cadence, and how much ops work you want to own.

This post is a decision framework, not a sales pitch. It favors Agent37 for some situations and self-hosting for others. For most teams, the honest answer turns out to be "both."

Key Takeaways

• Agent37 is Nous's hosted Hermes service from $3.99/mo; self-host costs $4-8/mo for a VPS plus your ops time.
• Price is rarely the deciding factor; data locality, compliance, and update cadence usually are.
• Hobbyists and solo developers: Agent37 wins on simplicity.
• Small teams handling customer data or secrets: self-host wins on control.
• Large teams often run both — Agent37 for dev/experimentation, self-host for production workloads with secrets.
• Self-hosting trades monthly dollars for weekly-ish ops attention (updates, monitoring, backups).
• Both options use the same Hermes binary; skills, memory format, MCP configs, and messaging gateways are portable between them.

What You Actually Get With Each Option

Agent37 is a managed daemon. Nous runs hermes-agent on their infrastructure, exposes the chat UI, handles updates, and bills you monthly. You point your Claude credentials (or use Anthropic keys managed in their dashboard), configure gateways, and get going. You do not SSH into anything. Updates happen on Nous's cadence.

Self-hosted is you installing the daemon somewhere you control. Typically:

• A $4-8/mo VPS (Hetzner, DigitalOcean, Vultr).
• Modal or Daytona for ephemeral or serverless variants.
• Your own Kubernetes cluster for teams already running infrastructure.

You run the install script, configure, update, and monitor. The binary is the same; only the operational context differs.

The Decision Axes

Cost at Scale

At small scale, prices are a wash. At larger team scale, a single $20/mo VPS can serve an entire engineering team if you do not need heavy parallelism; Agent37 tiers scale per-user or per-use. Run the numbers for your team shape.

Data Locality

This is usually the first axis that matters.

• Memory written to ~/.hermes/ on Agent37 lives on Nous's infrastructure. They document their handling, but the data is not on your machines.
• Memory written to ~/.hermes/ on your VPS lives on your VPS. You control retention, encryption at rest, and backups.

If your Hermes is reading production database rows, customer emails, or internal docs, the data locality axis strongly prefers self-hosting. If it is summarizing your public GitHub activity, Agent37 is fine.

Compliance

For regulated environments (HIPAA, SOC 2, data-residency requirements):

• Agent37 may or may not map to your framework — check their compliance documentation current to your date.
• Self-hosting in the region and infrastructure you already use for production is the simpler compliance story. If your production is in eu-west on a specific provider, Hermes goes there too.

This is not a place to cut corners. If compliance is on the table, the answer is almost always self-host inside the environment already covered by your audits.

Update Cadence

• Agent37: Nous updates the daemon on their schedule. You get new features automatically. You also inherit whatever changes land, including breaking ones.
• Self-host: you decide when to run hermes upgrade. You can pin a version until you have tested it.

If you are running Hermes for production incident response, pinning matters. If you are tinkering, auto-updates are a feature.

Ops Capacity

Honest self-assessment:

• Do you have someone who already manages a VPS? Self-host is easy; add Hermes to their list.
• Are you a solo developer who does not want a Linux box to babysit? Agent37 is the answer.
• Do you want to deploy on Modal or Daytona for serverless economics? Self-host; Agent37 does not (at time of writing) offer ephemeral-runtime tiers.

The Decision Matrix

                         | Agent37  | Self-host VPS
-------------------------+----------+--------------
Solo dev, hobby use      |   YES    |   optional
Solo dev, prefers UI     |   YES    |   no
Team of 2-5, no secrets  |   YES    |   optional
Team of 2-5, w/ secrets  |   no     |   YES
Large team (10+)         |   dev    |   prod
Regulated environment    |   no     |   YES
Need ephemeral runtime   |   no     |   YES (Modal)
Want version pinning     |   no     |   YES
No ops capacity          |   YES    |   no
Cost-sensitive at scale  |   varies |   usually YES

Recommended Patterns

Hobbyist / Solo Developer

Agent37. Sign up, point it at your Anthropic key, connect the messaging gateways you use, and done. You get the full Hermes experience without a VPS. If you outgrow it, migration is straightforward because the skills and memory formats are identical.

Small Team with Secrets

Self-host on a $5/mo VPS in the same region as your production systems. Run under a dedicated user, encrypt the ~/.hermes/ directory at rest (VPS-level LUKS or provider-managed encryption), and back up the memory store nightly. Add TLS-fronted gateways only for the interfaces you actually need.

See installing Hermes Agent on a 5 dollar VPS for the concrete setup.

Large Team

Run both. Agent37 for experimentation and dev tooling (engineers try new skills without asking ops for a VPS). Self-hosted in your production environment for the agent that does incident response, reads production logs, and holds sensitive memory. Skills flow between them because the format is the same.

Example split:

• Agent37 tier: engineers' personal Hermes instances, no production-data access.
• Self-hosted production Hermes: behind your VPC, production credentials, PagerDuty webhook, Slack gateway only on internal channels.

Migration Between the Two

Migration cost is small because the underlying pieces are portable:

• Skills: markdown files; move the directory.
• Memory: FTS5-indexed markdown in ~/.hermes/memory/; rsync the folder.
• MCP servers: same config.yaml format; port with minor path edits.
• Messaging gateways: reconfigure tokens against the new host.

The only gotcha is long-running webhook subscriptions (gateways that register a callback URL). Update those URLs at the provider when you move.

One More Axis: Psychological Overhead

Less tangible, but real: some engineers sleep better knowing they can SSH into the box running their agent. Others sleep better knowing they never have to. That is a valid criterion. Pick the option that reduces your anxiety surface; this is a tool you will interact with daily.

For the broader question of Hermes versus Claude Code, see Hermes vs Claude Code: when to use which. For the ecosystem context, see Hermes vs OpenClaw: two approaches to persistent AI agents.

Sources

• GitHub: NousResearch/hermes-agent
• Hermes docs: hermes-agent.nousresearch.com/docs/
• Agent37 pricing page (Nous Research)
• Related: Installing Hermes Agent on a 5 dollar VPS
• Related: Hermes vs OpenClaw: two approaches to persistent AI agents
• Related: Hermes vs Claude Code: when to use which