Summary: Nous Research launched Hermes Agent on February 25, 2026 as an MIT-licensed, self-hosted persistent agent that keeps memory on the user’s hardware and accumulates reusable skills. It led OpenRouter in cumulative token use by mid-August, but the article stresses that its always-on runtime and single-channel ranking make comparison with Claude Code unreliable. Desktop beta, funding talks at a $1.5 billion valuation, and local-skill portability define its adoption and governance tradeoffs.
If the 2026 agent market had one defining question, it was whose machine holds the memory and whose machine does the work. Nous Research’s Hermes Agent staked out the most emphatic answer: open source under the MIT license, self-hosted, with memory stored on hardware you own. It launched on February 25, 2026, and by mid-August it stood first on OpenRouter by cumulative token consumption.
What it is
Hermes Agent is best understood as a persistent personal agent. It lives on your own machine, holds onto your projects and preferences from one session to the next, and since it is a process that runs around the clock rather than a command you fire and forget, it reads as an assistant instead of a one-shot tool.
You do not open a web page to reach it; you use the messengers already on your phone. Telegram, Discord, Slack, WhatsApp, Signal, and Email all connect, and a CLI is available too. Scheduling is expressed in natural language, so jobs like a morning report, a nightly backup, or a weekly audit run unattended in the background. When a job has to be divided, it opens isolated subagents, each with a conversation and a terminal of its own, coordinated over Python RPC.
What sets it apart most is that skills accumulate on their own: a problem it has worked through gets distilled into a reusable skill, and the layer thickens the longer you use it. Officially, “it learns your projects, auto-generates skills.”
Desktop beta: the terminal, moved into a window
An official Hermes Desktop beta has been shipping since June 2, offering one-click installers for macOS and Windows and a terminal install for Linux. The stated position is “mirror, not replace”: whatever the CLI can do, the app can do, with file browsing, previews, and a plugin system for tools, hooks, themes, and integrations layered on top.
The base app requires no account. Always-on cloud agents and model discounts arrive only once you sign in to Nous Portal. That boundary is a product decision in itself: the core stays free and open, and revenue attaches to hosted execution and model resale.
Number one in usage, if you read it right
Third-party numbers from mid-August came with a headline built to grab attention: by cumulative token consumption on OpenRouter, Hermes Agent leads at 35.7 trillion, ahead of Claude Code at 8.53 trillion, Kilo Code at 7.36 trillion, and the early-year viral hit OpenClaw, fallen to 4.4 trillion. Hermes first took the daily top spot on May 6, burning 271 billion tokens that day.
The same article, though, supplies three caveats of its own that deserve to be carried over exactly as written. First, Hermes is a 24/7 persistent runtime while Claude Code is session-based, so the raw token totals are not comparable by construction. Second, most Claude Code, Cursor, and Copilot traffic never passes through OpenRouter, so the ranking captures exactly one channel. Third, the table keeps moving — OpenRouter itself is being acquired by Stripe for more than $7 billion.
What the numbers actually establish is narrower than the headline: a large population is willing to let a self-hosted open-source agent burn tokens around the clock. They do not establish that Hermes beat Claude Code.
Money and ecosystem
TechCrunch reported on July 13 that Nous Research is in talks for a new round at a $1.5 billion valuation, led by Robot Ventures with USV taking part. For a company that made its name on open models, that number is pinned straight to Hermes Agent’s adoption curve.
The timing is worth noticing too. Hermes arrived as one of the challengers after OpenClaw’s viral run, and OpenClaw has spent the half-year since sliding: 138 CVEs accumulated in five months, one of them rated CVSS 9.9, and its founder joined OpenAI in February. A good share of the market’s trust in open-source agents is borrowed from a competitor’s security incidents — the same script we walked through in OpenClaw’s two renames and the security concerns.
The design bets that actually matter
Keeping memory local buys data sovereignty and charges you the risk: no cloud vendor patches for you, backs you up, or absorbs a breach on your behalf. Letting the agent write its own skills buys compounding convenience and charges you at review time: a bad procedure, once fossilized as a skill file, gets repeated faithfully ever after.
For teams doing procurement, the portable skill asset is the part I would copy. Models and frameworks can both be swapped out; a procedure distilled into an open-format skill file, though, is an asset that travels with you. The industry is moving the same way — LM Studio’s Bionic and Meta’s open-weight Muse Glimmer are saying the same thing: the agent belongs on your machine.