ooligo

Hermes Agent

ai-agent-platform ai-agents · self-hosted · ai-assistant · open-source
AI-NATIVE MCP API FREEMIUM
RevOpsLegal OpsRecruiting & TA
7.9 /10

What it is

Hermes Agent is Nous Research’s open-source, self-hosted AI agent, released under the MIT license in February 2026. You install it on a laptop or a small VPS, point it at a model provider, and talk to it from a terminal or from the chat apps your team already opens. One gateway process serves Telegram, Discord, Slack, WhatsApp, Signal and email, and the docs list more than 20 platforms in total, including Microsoft Teams, Google Chat, Matrix and Mattermost. The GitHub repository sat at roughly 263,000 stars when we checked on 13 November 2026.

What Nous sells as the difference is the learning loop. After a turn, the agent can write or patch its own skills: folders under ~/.hermes/skills/<category>/<name>/SKILL.md, in the same agentskills.io format Claude Skills use. It also keeps two small memory files, MEMORY.md and USER.md, that load into every session. The agent you run in month three has instructions it wrote for itself, which is the point and also the main governance question.

The company changed shape in October 2026. On 7 October Nous announced a $90M round with Nvidia, Samsung, Microsoft’s M12, Union Square Ventures, Y Combinator and Menlo Ventures participating, to build paid business editions. The $1.5B valuation comes from the Wall Street Journal; Nous did not state one. Usage figures (over 24 million clones, about 2.5% of global token usage) are Nous’s own estimates and should be read as such.

Three ways to buy it

  • Self-hosted, $0. Install with curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash on Linux, macOS or WSL2, or the PowerShell installer on Windows, then hermes gateway setup to connect channels. Model spend goes to whichever provider you attach: the Nous Portal, OpenRouter, OpenAI, or any endpoint you configure.
  • Hermes Cloud. Nous hosts the agent for you on a metered instance billed daily from portal credit: compute by the second while running, storage only while stopped, inference and paid tools on top. You need $2 in credit or a subscription to deploy. Portal plans are Free (free models only), Plus at $20/month (4 GB RAM, 25 GB disk), Super at $100/month and Ultra at $200/month, and each paid plan includes one Cloud agent.
  • Hermes for Business. A shared credit balance with per-member spend caps, admin-managed roles and API keys, a team skill library, and a team gateway for Discord, Telegram, Slack and WhatsApp, run in an isolated tenant per team on Nous infrastructure. Signup is self-serve, but no price is published. SSO, SLAs and on-prem or private-cloud deployment sit in a separate Enterprise tier sold through sales.

What it costs in practice

For one operator, the realistic monthly bill is the model, not the software. Self-hosting adds a server you size yourself, since Nous publishes no hardware floor; Hermes Cloud’s Plus plan at $20 removes the server but not the inference line, which scales with every turn and every background skill review. The business edition is where the number becomes unknowable from the outside: with no price list, treat any figure you hear as a quote to verify in writing, and model it against Claude Cowork at $25 per seat per month on Claude Team ($20 billed annually), which also bundles the model.

Best for

  • RevOps, Legal Ops or Recruiting operators with a technical owner who want one always-on agent in Slack or Telegram that keeps procedures it learns (a weekly pipeline-hygiene sweep, a contract-intake triage, a candidate-digest skill) and runs on the model they already pay for.
  • Small teams testing Hermes for Business where a shared credit pool with per-person caps is the control the finance lead actually asked for.

Alternatives, and when to pick them instead

  • OpenClaw: the closest peer and, with Claude Cowork, the agent most ops teams already have on the shortlist; since July 2026 it is run by a non-profit foundation. Pick it when you want governance continuity over self-written skills. It has no vendor selling a paid edition, so there is no upsell pressure and no support contract either.
  • QM: pick it when the requirement is multiplayer, with per-channel agents and per-scope permissions and audit logs inside your own AWS or Fly.io account. Hermes scopes memory per profile, not per Slack channel.
  • Claude Cowork: pick it when nobody will own a daemon and the company already pays for Claude. No host, no gateway, scheduled tasks in Anthropic’s cloud, and the procurement paperwork already exists.
  • Hermes is itself the fastest-growing entrant here: a February 2026 release that reached roughly 263,000 GitHub stars in nine months. That is a reason to test it and a reason to pin a version rather than run the installer’s latest build in production.
  • Manus: pick it for one-off research and browsing tasks a non-engineer launches, not for a resident agent that accumulates procedures.

Watch-outs

  • Self-written skills change behavior without a review. By default the agent creates and patches skills after turns, and memory writes are not gated either. Guard: set skills.write_approval: true (writes stage in ~/.hermes/pending/skills/ until approved) and memory.write_approval: true, and keep ~/.hermes/skills/ under git so every change has a diff.
  • Container backends skip the dangerous-command check. The docs say that in Docker, Modal and the other sandboxed backends “the container itself is the security boundary”, so command screening is off; on the local backend, the file write guards are described as defense in depth, not a hard boundary. Guard: never mount credentials or a shared drive into the sandbox, and keep approvals.mode at manual for any agent that reads email or web pages from outside the company.
  • The default approver list is everyone. approvals.approvers defaults to ["*"], so anyone allowed on the gateway can approve a risky command. Guard: name approvers explicitly, use the gateway allowlist with DM pairing, and run one profile per trust boundary rather than one shared team agent.
  • Memory is small and fails loudly. MEMORY.md caps at 2,200 characters and USER.md at 1,375; when full, writes fail until the agent consolidates. Guard: put durable reference material in skills or an external memory plugin (Honcho, Mem0 and others are supported), not in core memory.
  • The business edition is new and unpriced. No published price, no audit log on the product page, and SSO only in Enterprise. Guard: for regulated data, require the DPA, sub-processor list and a written price before moving anything off the self-hosted install.