Hermes Agent v2.0! Huge New Updates: WebUI, Qwen 3.6 Plus FREE, Computer Use, & More!

WorldofAI · 11m 07s · Watch on YouTube · 22 sources

Decision Card

Effort: One evening (~2–3 hours) to try the core of it — install Hermes Agent from github.com/NousResearch/hermes-agent, create a free Nous Portal account for the temporarily-free Qwen 3.6 Plus, run hermes update + hermes dashboard to see the Kanban board; computer use and LightPanda are separate ~30-minute add-ons each.

Honest take: The video gets the driver name wrong (“powered by Cocoa” — it is actually the open-source cua-driver) and its “macOS-only for now” claim is already stale: the official docs list background computer use on macOS, Windows, and Linux as a supported feature, not an “early preview.” The opening leaderboard claim (“climbing above OpenClaw, Kilo Code…”) is asserted without naming any leaderboard, and a two-minute TinyFish sponsor read is spliced mid-sentence into the computer-use segment, so it’s easy to mistake sponsor features for Hermes features.

Concrete next steps:

  • Computer usetry: run hermes computer-use install (or hermes tools), then hermes computer-use doctor for permission checks; needs any vision/tool-capable model (~30 min). Skip if you won’t grant an agent OS-level accessibility permissions.
  • Qwen 3.6 Plus free tiertry: free account at portal.nousresearch.com (~10 min); it’s “free for a limited time,” so test it before deciding whether Portal’s subscription model beats a plain OpenRouter key. Skip if you already have a satisfactory frontier-model subscription.
  • LightPanda browser backendtry: needs Hermes v0.13.0+ and LightPanda on PATH; setup guide at lightpanda.io/docs/open-source/guides/use-hermes (~30 min). Skip if your agent rarely browses — Chrome fallback already covers everything.
  • Kanban board + /goal commandtry: hermes update, then the init command from the Kanban tutorial, then hermes dashboard (~15 min). Skip if you run single short-lived sessions; the board’s value is multi-agent, long-running work.
  • TinyFish free search/fetch APIstry independently of Hermes: free per the announcement, docs at docs.tinyfish.ai (~20 min). Skip if you already pay for Tavily/Exa and are happy — note this is the sponsor segment, so treat the latency claims skeptically.
  • Voice cloning / TTS extrasskip unless you specifically want spoken agent output; the video covers it in one sentence with no setup detail.

TL;DR

WorldofAI rounds up a week of Hermes Agent updates: background computer use that works with any vision-capable model, Alibaba’s Qwen 3.6 Plus temporarily free on Nous Portal, LightPanda as an integrated open-source browser backend with Chrome fallback, a revamped Kanban web dashboard for multi-agent orchestration, and a /goal command for long-running autonomous objectives. The information is broadly accurate but shallow — feature names and platform availability are partly wrong or outdated versus the official docs, so verify against the docs before adopting.

Key Points

  • Hermes Agent is an open-source (MIT) autonomous agent from Nous Research designed to run 24/7 on your own infrastructure while accumulating long-term memory, reusable skills, and a model of its user. 00:26
  • New computer-use support lets Hermes control the machine in the background — your cursor doesn’t move and keyboard focus doesn’t change while the agent clicks, types, and scrolls alongside you. 01:23
  • The video claims computer use is macOS-only for now with Windows/Linux “soon” — the official docs now list all three platforms as supported. 03:18
  • Unlike most computer-use systems, it works with any tool-capable vision model — Claude, GPT, Gemini, or local open-source models via vLLM — rather than only frontier models with dedicated computer-use modes. 03:43
  • Enable it with hermes computer-use install from the CLI, or interactively via hermes tools. 04:00
  • Alibaba’s Qwen 3.6 Plus — a 1M-token-context model strong at web development, long-horizon and agentic tasks — is free for a limited time on Nous Portal. 04:47
  • Nous Portal is built on top of OpenRouter, so you get OpenRouter’s model routing plus exclusive free models, discounts, and bundled tool usage tuned for Hermes workflows. 05:30
  • LightPanda, an open-source browser built for machine workloads, is now an integrated browser backend in Hermes with automatic Chrome fallback when a page needs features LightPanda lacks. 06:06
  • The Kanban board got a major update: unlimited boards/projects, multi-agent task management, and project-update subscriptions pushed to your home channel through the configured gateway messenger; start it via hermes update then the dashboard command. 06:53
  • The new /goal command runs a long-lived objective loop — continuously plan, execute, review, and retry on failure, managing subtasks until the objective completes — tracked through the Kanban board. 09:25

Notable Quotes

“It’s built by Nous Research under the MIT license, and it is an agent that can run 24/7 on your own infrastructure while building long-term memory, reusable skills, and even a deeper understanding of itself as you use it more.” 00:33

“Hermes does not fully take over your PC. You can still continue using your keyboard, mouse, apps, and workflows normally like you would usually while the AI agent operates alongside you simultaneously.” 01:29

“This essentially turns Hermes into an even more of a persistent autonomous workspace with multi-agent operating environments rather than just a standalone AI tool.” 07:13

Verified Claims

  1. Hermes Agent is open source under the MIT license, built by Nous Research. 00:30 — Sources: NousResearch/hermes-agent LICENSE, GitHub repo. Verdict: Confirmed.

  2. Computer use runs in the background — cursor doesn’t move, keyboard focus doesn’t change, macOS Spaces don’t switch. 03:34 — Source: Computer Use docs, which state exactly this, powered by the open-source cua-driver (using AX on macOS, UIAutomation on Windows, AT-SPI on Linux). Verdict: Confirmed — though the video’s “powered by Cocoa” is a mangling of “cua.”

  3. Computer use is macOS-only at the moment, with Windows and Linux coming soon. 03:18 — Source: the official docs list macOS, Windows, and Linux as all supported. Verdict: Disputed (likely true at recording, outdated now).

  4. It works with any tool-capable model — Claude, GPT, Gemini, and local models via vLLM — not just frontier models with special computer-use modes. 03:43 — Source: Computer Use docs (“any tool-capable model,” including local OpenAI-compatible endpoints). Verdict: Confirmed.

  5. Qwen 3.6 Plus is available on Nous Portal, free for a limited time. 04:47 — Sources: Qwen’s announcement on X, Nous Research on X. Verdict: Confirmed.

  6. Qwen 3.6 Plus supports a 1 million token context window. 05:15 — Sources: BuildFastWithAI review, MindStudio overview. Verdict: Confirmed.

  7. Nous Portal is built on top of OpenRouter, giving OpenRouter routing benefits plus exclusive free models and bundled tools. 05:30 — Sources: Nous Portal integration docs, Nous Portal guide (“routing happens through OpenRouter under the hood… billed against your Nous subscription”). Verdict: Confirmed.

  8. LightPanda is integrated as a browser backend in Hermes, settable as default with automatic Chrome fallback. 06:28 — Sources: LightPanda announcement blog, Hermes PR #7144. Requires Hermes v0.13.0+; Chrome still handles unsupported actions like screenshots. Verdict: Confirmed.

  9. TinyFish’s web search and fetch APIs are now completely free with no caps. 02:18 — Source: TinyFish blog announcement. Free with “generous rate limits” — so “no caps” is an overstatement; browser and agent endpoints remain metered. Verdict: Confirmed with caveat (sponsor claim). The “under 500 ms” latency figure was not independently verifiable.

  10. The /goal command runs a continuous plan–execute–review–retry loop until the objective completes, tracked via the Kanban board. 09:25 — Sources: Kanban docs, Hermes Kanban deep-dive (a judge model evaluates each turn against the goal and feeds continuation prompts until met). Verdict: Confirmed.

The opening claim that Hermes is “climbing above tools like OpenClaw and Kilo Code across many daily-use AI agent leaderboards” names no leaderboard and could not be verified. Verdict: Inconclusive.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. How does cua-driver’s reliance on private macOS SkyLight SPIs (which Apple can break in any OS update) affect the durability of background computer use, and what is the failure mode when those APIs change?
  2. What does Nous Portal’s pricing look like after the Qwen 3.6 Plus free window ends, and does the bundled-subscription model actually beat a direct OpenRouter key for typical Hermes usage?
  3. What fraction of real-world agent browsing does LightPanda handle natively versus falling back to Chrome, and how much memory/cold-start time does it actually save in long-running deployments?

Sources