AI & Automation / AI Workflows

Hermes Agent v2.0: WebUI, Computer Use, Kanban, and /goal Guide (05/13/26)

Use this guide to understand, install, and safely evaluate the Hermes Agent v2.0-style workflow shown in the WorldofAI video: background computer use on macOS, free Nous Portal model routing, Lightpanda-backed browser automation, Kanban multi-agent boards, and persistent /goal loops.

Outcome Update Hermes, enable key tools, and run a controlled test.
Audience Intermediate AI-agent users and homelab operators.
Primary OS macOS for Computer Use; core Hermes works on Linux/macOS/WSL2.
Source limits YouTube auto-transcript + metadata, checked against official docs.

1. What the video covers

The video presents Hermes Agent as an open-source, persistent, self-improving agent from Nous Research. The main claim is not “another chatbot,” but an always-available agent that can remember work, reuse skills, run scheduled automations, delegate to agents, and operate through CLI, WebUI, and messaging channels.

Major updates highlighted

  • Background Computer Use on macOS through cua-driver.
  • Nous Portal model routing, with the video highlighting Qwen/Qwen3.6-Plus availability as free for a limited time.
  • Lightpanda-style browser backend and improved autonomous web tasks.
  • Kanban boards for durable multi-agent coordination.
  • Persistent /goal mode for long-running objectives.

Recommended evaluation path

  1. Install or update Hermes.
  2. Pick a model/provider you can afford and trust.
  3. Enable only one advanced capability at a time.
  4. Run small, reversible tasks first.
  5. Use Kanban and /goal only after basic CLI sessions are stable.
Source note: the video transcript appears to be automatically generated, so names like “Nous,” “Qwen,” “Kanban,” and “Claude” were normalized where the transcript produced obvious speech-to-text errors. Commands and platform limits were checked against the official Hermes docs where available.

2. Prerequisites

  • A machine for Hermes: macOS, Linux, WSL2, Termux, or native Windows beta depending on your tolerance for rough edges.
  • Python and shell access. The official installer handles most dependencies.
  • At least one model provider configured, such as Nous Portal, OpenRouter, OpenAI, Anthropic, Google, local vLLM, or LM Studio.
  • For Computer Use: macOS, Accessibility permission, Screen Recording permission, and a vision-capable model for best results.
  • For remote access: SSH tunnel, Tailscale, or another private network approach. Do not expose a control panel to the public internet without authentication.
Security posture: Computer Use can click, type, drag, and scroll. Keep approvals manual until you have watched it succeed on boring tasks. The exciting demo is the easy part; not letting a robot click something cursed is the adult supervision bit.

3. Install or update Hermes

If Hermes is not installed, use the official installer for your platform. On Linux, macOS, WSL2, or Termux:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.zshrc  # or source ~/.bashrc
hermes

If Hermes is already installed, update before testing the features from the video:

hermes update
hermes doctor

Then select or verify a model provider:

hermes model
hermes tools

For migrations from OpenClaw, Hermes documentation describes hermes claw migrate, including dry-run and preset options. Use a dry run before importing secrets or workspace data.

4. Enable background Computer Use on macOS

The video’s most eye-catching update is background computer use. Official Hermes docs describe this as a macOS-only feature that lets Hermes click, type, scroll, and drag without moving your cursor or stealing keyboard focus.

Install path

hermes computer-use install
hermes computer-use status

Or enable it interactively:

hermes tools
# choose: Computer Use (macOS) → cua-driver (background)

Grant macOS permissions

  1. Open System Settings → Privacy & Security → Accessibility.
  2. Allow your terminal app, Hermes app, or whichever host launches Hermes.
  3. Open System Settings → Privacy & Security → Screen Recording.
  4. Allow the same app.
  5. Restart the terminal/app if permissions do not take effect immediately.

Start a session with the toolset

hermes -t computer_use chat

Or add computer_use to your enabled toolsets in ~/.hermes/config.yaml.

Safe first test: ask Hermes to capture a screenshot of a non-sensitive app and describe the visible UI. Then test one click in a harmless app. Do not start with banking, admin panels, or password entry. Humanity has suffered enough from “let’s just try it in prod.”
CapabilityWhat to expectConstraint
CaptureScreenshot or accessibility tree with selectable elements.Screenshots cost tokens; keep tasks short.
Click/type/scrollAgent acts in the target process while your cursor stays put.Destructive actions should require approval.
Vision modelsClaude, GPT, Gemini, OpenRouter vision models, or local VLMs can interpret screenshots.Text-only models degrade to accessibility-tree operation.
Platform supportmacOS only for this background driver.Use browser automation for cross-platform web tasks.

5. Configure models and Nous Portal

The video highlights Qwen/Qwen3.6-Plus as available through Nous Portal and free for a limited time. Treat “free” and “limited time” as changeable service terms: verify the current model list and pricing inside Nous Portal before building recurring workflows around it.

  1. Create or open a Nous Portal account.
  2. Copy the API key or provider settings required by Hermes.
  3. Run hermes model and configure Nous Portal or your preferred provider.
  4. Inside a session, switch models with /model when needed.
Model-selection rule of thumb: use a strong vision-capable model for Computer Use, a fast inexpensive model for judge/helper tasks, and your best reasoning model for long-running /goal or orchestration work.

6. Browser automation and Lightpanda-style backend

The video describes improved browser automation, including a Lightpanda browser backend and automatic Chrome fallback. The practical reason this matters: agent web tasks often fail when a browser backend is too heavy, too brittle, or too different from real web behavior.

Use browser automation when the task is web-only and does not require native desktop UI. Use Computer Use only when Hermes must interact with a real local app, system dialog, or desktop workflow.

NeedBest toolWhy
Read a website or click through a login flowBrowser toolsetMore portable and easier to constrain.
Operate Mail, Finder, native apps, or desktop UIComputer UseCan target macOS apps directly.
Long research with many pagesBrowser/search/fetch pipelineCheaper and less risky than full desktop control.

7. Set up Kanban and the WebUI/dashboard workflow

The video frames Kanban as the big step from “single agent chat” to “observable multi-agent workspace.” Official docs define Hermes Kanban as a durable task board backed by SQLite, shared across Hermes profiles, with tasks, comments, dependencies, run history, and worker handoffs.

Initialize and open the dashboard

hermes kanban init   # optional; first kanban command auto-inits
hermes dashboard     # opens the dashboard, then click Kanban

Expected board columns are Triage, Todo, Ready, In progress, Blocked, and Done.

Create a small multi-agent test

hermes kanban create "Draft a one-page summary of Hermes v2 updates"   --assignee writer   --body "Use official docs and the video notes. Include setup caveats."

hermes kanban list

For dependency chains, create parent/child tasks. The dispatcher promotes children only after parents are done, which is exactly the sort of unglamorous bookkeeping humans hate and agents are weirdly good at.

When to use Kanban instead of a normal subagent

  • Work needs to survive restarts.
  • A human may need to comment, unblock, or review.
  • Several named profiles should collaborate over time.
  • You need audit history, run summaries, or repeatable handoffs.

8. Use /goal for long-running objectives

/goal gives Hermes a standing objective that continues across turns until a judge model decides the goal is done, blocked, paused, or out of turn budget. This is useful when you would otherwise keep saying “continue” like a disappointed project manager.

/goal Fix every failing test in tests/hermes_cli/ and verify the relevant test command passes

Helpful control commands:

CommandUse
/goal statusSee the active goal and progress.
/goal pauseStop the continuation loop without deleting the goal.
/goal resumeResume the loop, typically resetting the turn counter.
/goal clearRemove the standing goal.
Use a budget: the official docs describe a default continuation budget of 20 turns. Keep that or lower it while testing; autonomous loops are delightful right up until they autonomously create a mess.

9. Success checks

  1. hermes doctor completes without serious errors.
  2. hermes model shows your intended provider and model.
  3. hermes computer-use status reports the driver installed on macOS.
  4. A computer_use capture works in a harmless app.
  5. hermes dashboard opens and the Kanban tab loads.
  6. hermes kanban create ... creates a visible card.
  7. /goal status reports the active goal after you set one.

10. Troubleshooting

ProblemLikely causeFix
computer_use backend unavailablecua-driver is not installed or not on path.Run hermes computer-use install, then hermes computer-use status.
Capture works but clicks do nothingmacOS permissions or a hidden modal is blocking interaction.Check Accessibility and Screen Recording permissions; capture again after dismissing modals.
Element numbers stop matchingSOM indices are stale after UI changes.Re-capture before the next action.
WebUI is reachable from the internetHost binding or port forwarding is too open.Bind to 127.0.0.1, use SSH/Tailscale, and set a strong password if exposed.
Kanban worker never startsGateway/dispatcher not running, profile missing, or task blocked.Start the gateway/dashboard, verify assignee profile names, inspect the task comments/run history.
/goal keeps looping after “good enough”Judge model is conservative or final response did not explicitly confirm completion.Ask for status, pause it, or finish with a clear completion statement.

11. Sources