Paperclip AI + Hermes Agent: How to Run Hermes as an Employee in Paperclip

Pairing Paperclip AI with Hermes Agent gives you something neither project offers alone: an agent that gets better at its job every week, inside a company structure that keeps it accountable. Hermes, built by Nous Research, learns skills from its own experience. Paperclip hires it, gives it a role, caps its spending and logs every decision. This post explains what Hermes is, how the Hermes adapter fits into Paperclip, which roles suit a learning agent, and where the combination still falls short. For the platform fundamentals, read our full Paperclip AI guide first.

What Hermes Agent is

Hermes Agent is an open-source autonomous assistant released by Nous Research in early 2026, in the same few weeks that OpenClaw and Paperclip appeared. It runs as a long-lived process with its own tool set, a memory store and a skills directory. The feature that sets it apart is the learning loop. When Hermes completes a task, it can distill what worked into a reusable skill file. The next time a similar task arrives, it loads that skill instead of reasoning from scratch. Over time a Hermes instance that handles your weekly reporting becomes noticeably faster and more consistent at weekly reporting.

Like OpenClaw, Hermes is a single agent. It does not know about other agents, has no org chart, no ticket queue and no budget enforcement. A Hermes instance left alone will happily keep improving at whatever it is asked to do, including things you never intended it to spend money on. That is exactly the kind of agent that benefits from a control plane.

Why a learning agent needs governance

Paperclip's pitch is that agents should be managed like employees rather than scripted like workflows. That framing matters more for Hermes than for any other adapter, for three reasons.

• Skills drift. A self-improving agent changes its own behavior. Paperclip's immutable audit log shows you exactly which skill was used on which ticket, so when output quality shifts you can trace it to the skill that changed.
• Cost compounds. Learning takes tokens. Paperclip's per-agent monthly budget with automatic pause at 100 percent means the learning loop cannot become a runaway bill.
• Scope creep. Goal alignment in Paperclip ties every ticket back to a company goal. A Hermes employee gets work that traces to a mission, not an open-ended prompt, which keeps its skill library focused on your business.

How the Hermes adapter works

The Hermes adapter follows the same contract as every other Paperclip adapter. Paperclip sends a heartbeat containing the agent identity, the checked-out ticket, the goal ancestry and the role's permissions. Hermes receives it as a task, runs its normal loop including any skill retrieval, and returns the result along with tool-call traces and token usage. Paperclip writes all of it to the ticket, charges the budget and advances the ticket to review or done.

One detail worth understanding: Hermes keeps its memory and skills on its own side. Paperclip does not store them. If you terminate a Hermes employee in Paperclip and spin up a new one pointing at the same Hermes instance, the learned skills are still there. If you delete the Hermes instance, they are gone. Back up the skills directory the way you would back up any other business asset.

Setup, step by step

Field labels vary by release (v2026.916.1 is current as of September 2026), so these steps describe intent rather than exact UI paths.

1. Install Paperclip. The fastest route is npx paperclipai onboard --yes, which brings up the dashboard at http://localhost:3100 with an embedded PostgreSQL. For a permanent server, follow our installation guide.
2. Install and start Hermes. Configure its model provider and confirm it responds to a direct task before involving Paperclip. Note the endpoint and credentials it exposes.
3. Create the role. In the Paperclip org chart, add a role with a specific job description and pick which skills the role may use. Hermes-authored skills can be allowed or restricted at this level.
4. Add the agent with the Hermes adapter. Point it at the Hermes endpoint, store the credential in Paperclip's encrypted secrets, and set the heartbeat schedule.
5. Set the budget and an approval gate. A modest monthly cap plus an approval stage on any customer-facing or money-moving output. You can loosen both once the skill library has matured.
6. Run a training ticket. Give Hermes three or four representative tasks in its first week. Watch the activity log for the skills it creates. Prune anything that looks wrong early, because Hermes will reuse it.

Roles where Hermes shines

Role	Why Hermes fits	Suggested heartbeat
Reporting Analyst	Same report shape every week; skills converge quickly	Weekly plus task-triggered
Data Cleanup Specialist	Repetitive transformations with edge cases worth remembering	Daily
Lead Researcher	Learns which sources are reliable for your niche	Every few hours
Content Repurposer	Style rules become skills after a handful of corrections	Task-triggered
QA Reviewer	Accumulates a checklist of past defects	Task-triggered on PR or draft

Roles built around live conversation, such as customer support over Telegram, are usually a better fit for OpenClaw because of its channel coverage. See our OpenClaw integration guide for that side.

The OpenClaw + Hermes + Paperclip stack

The pattern that has become common in 2026 looks like this: OpenClaw employees on the front line handling messaging channels, Hermes employees in the back office doing recurring analytical work, and a Claude Code or Codex agent as a manager that reads the backlog and delegates. Paperclip sits over all of it as the company: org chart, goals, tickets, budgets, approvals and audit. Each agent type does what it is best at, and none of them needs to know the others exist. We describe how the manager layer works in how Paperclip AI orchestration works.

Limitations to know before you commit

• Two moving targets. Both projects ship frequently. An update on either side can change the adapter contract. Pin versions on production and update on a schedule.
• Skills are not reviewed by default. Hermes writes skills for itself. Unless you add an approval gate for skill changes, a bad skill can persist silently. Review the skills directory weekly early on.
• No shared memory across Hermes employees. Two Hermes instances do not share what they learned. If you want shared knowledge, put it in a company skill in Paperclip instead.
• Learning curve. Reviews of Paperclip in 2026 score it around 7.8 out of 10 and consistently cite a learning curve and some rough edges. Adding a self-modifying agent on top raises that bar. Budget a few weeks, not a few hours.

If you are still deciding between agents, our roundup of Paperclip AI alternatives compares Hermes, OpenClaw, Multica and the framework-style options side by side.

What the first week looks like

Expect the first few tickets to be slow and slightly off. Hermes has no skills for your business yet, so it reasons from scratch and asks for clarification more than you would like. By the end of the week the activity log usually shows two or three skills created, and ticket duration drops noticeably. Resist the urge to widen the role's scope during that week. Give it the same kind of task repeatedly, correct the output through the review stage, and let the skill library settle before you hand it anything new. That discipline is what separates a Hermes employee that becomes reliable from one that accumulates half-formed skills for six different jobs.

Frequently asked questions

Is Hermes Agent the same as the Hermes language models?

No. Nous Research publishes both a family of Hermes models and the Hermes Agent runtime. The agent can use those models, but it is a separate open-source project that runs as an autonomous assistant.

Does Paperclip store the skills Hermes learns?

No. Skills and memory live in the Hermes instance. Paperclip records which skills were used on each ticket in its audit log, but it does not own or back up the skill files.

Can I stop Hermes from learning new skills?

You can restrict which skills a role may use in Paperclip, and you can gate output behind approvals, but the learning loop itself is configured on the Hermes side. Consult the Hermes documentation for the setting.

Should my first Paperclip employee be Hermes or OpenClaw?

If the first job involves customers on messaging apps, OpenClaw. If it is a recurring internal task you want done better each week, Hermes. Many companies end up with both.

What does it cost?

Paperclip and Hermes are both free and open source. You pay for hosting and model usage, which Paperclip caps per agent. See our pricing breakdown.