↓ Skip to main content
tomrochette.com
  1. Agents/

Hermes Agent

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~797 words)
agent-curated 173 assistant-runtimes 10 fully-ai-generated 263 llm=glm-5.3 170 llm=glm-5.3-flash 178 nous-research 1 open-source 62 personal-assistants 6 research-note 148
Table of Contents

Hermes Agent is Nous Research’s MIT-licensed self-improving personal agent: a Python gateway process that lives in your chat channels, builds skills from its own experience, and runs on everything from a $5 VPS to serverless sandboxes, at about 244k stars the biggest runtime launch since OpenClaw. Facts below verified as of 2026-09-13.

Hermes’ bet is that the runtime winner is decided by the learning loop, not the channel list: an agent that curates its own memory, writes its own skills, and models you across sessions compounds while the others merely answer.

What it is
#

Install with a curl script (the installer bundles uv, Python 3.11, Node.js, ripgrep, ffmpeg, and an isolated Git Bash on Windows), and one gateway process serves Telegram, Discord, Slack, WhatsApp, Signal, and a CLI, with voice-memo transcription and conversation continuity across platforms. The learning loop is the differentiator: it creates skills autonomously after complex tasks, improves them during use, nudges itself to persist knowledge, searches its own past sessions (FTS5 with LLM summarization), and builds a user model through Honcho, compatible with the agentskills.io open standard. Any model goes in (Nous Portal, OpenRouter, OpenAI, custom endpoints) and switching is a hermes model command with no code changes. Execution spans seven backends (local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox), with serverless persistence that hibernates when idle, and a built-in cron scheduler delivers reports to any platform.

Status
#

Massive and fast-moving. As of 2026-09-13: 244,961 stars and 50,865 forks since creation on 2025-07-22, pushed today, MIT, and 42,428 open issues, a support surface bigger than most projects’ users. The 52-point launch thread landed June 2026, and a real tool ecosystem followed (a 28-point Grafana observability integration, a Fleet console for managing Dockerized Hermes agents, XMPP interop with OpenClaw). The governance record is the caution: a May 2026 thread documents Nous editing a GitHub issue to remove plagiarism claims about Hermes Agent (issue 10232), so origin claims around this project deserve independent reading.

Strengths
#

  • The learning loop (skills that write and improve themselves) is the most complete self-improvement story in the category.
  • Seven execution backends with serverless idle pricing solves the always-on cost problem that keeps personal agents expensive.
  • Model-agnostic to the point of indifference, with a one-command switch.
  • Nous Research’s model portfolio gives it a first-party provider path the -claw tools lack.

Cautions
#

  • 42,281 open issues is triage weather, not a queue; expect to read code when things break.
  • The plagiarism-claim edit is a stewardship red flag worth weighing before making it your memory keeper.
  • A self-curating memory compounds errors as efficiently as insights; the loop needs supervision, not just trust.
  • Windows installs bundle a lot of machinery (uv, Node, Git Bash), which is surface area to audit.

Pricing
#

The runtime is free and open source under MIT. Paid plans now exist through Nous Portal, and as of 2026-09-13 the prices are public on the official site: Free at $0, Plus at $20/month, Super at $100/month, and Ultra at $200/month, where each paid tier bundles monthly credits with a 10 percent bonus ($22, $110, and $220 respectively), 200+ models, and hosted tool use. Or bring your own provider keys, with serverless backends billed by the platform when idle-cheap.

Compared to
#

  • OpenClaw: the other giant, ecosystem-first; choose OpenClaw for channels and companions, Hermes for the learning loop and model freedom.
  • NanoClaw: the auditable minimalists’ pick; Hermes is the opposite trade, a lot of machinery you must trust.
  • Paperclip: if Hermes is an employee, Paperclip is the company that would manage several of them.

Bottom line
#

Recommended for engineers who want their assistant to get measurably better at their job and will supervise the loop. Not for minimal-footprint deployments or anyone uncomfortable with an agent that edits its own memory. The disagreeable claim I will defend: the channel-list era of this category is over and the memory-loop era decides the winner, and Hermes is a year ahead of the -claw family on exactly that axis.

Changes
#

  • 2026-08-27 - Created as an Assistant runtimes note after the control-planes alternatives scan.
  • 2026-09-09 - Added paid tiers (Free, Plus, Super, Ultra via Nous Portal), retiring the no-paid-tier claim.
  • 2026-09-12 - Paid tier prices went public on the official site, retiring the bot-wall claim.

See also
#

References
#