Hermes Workspace: What It Is, What It Isn't, and Which Path Fits Your Agent (2026)

Hermes Workspace, the on-disk workspace folder, and the official dashboard are three different things. Here is what each one is and how to run it always-on.

•

Published on

•

Hermes Workspace: What It Is, What It Isn't, and Which Path Fits Your Agent (2026)
Do not index
Hermes Workspace is an open-source web UI for the Hermes agent: a self-hosted command center with chat, a terminal, a memory editor, and multi-agent controls, built by the community rather than by Nous Research. The same phrase also gets used for two other things, the agent's on-disk working directory and the official built-in dashboard, and mixing the three up is how most setups go sideways. This guide untangles all three, then shows what each path costs to run always-on, including managed hosting on Agent37: a free sleeping personal agent to try, always-on Hermes from $3.99/mo.

What is Hermes Workspace, exactly?

Hermes Workspace is a community project, not an official Nous Research product. It lives at github.com/outsourc-e/hermes-workspace: an MIT-licensed web app with about 6,700 GitHub stars, built by developer outsourc-e outside Nous Research. Nous's own names are the Hermes agent, the built-in dashboard, the Hermes Desktop app, and its hosted product, Hermes Cloud. Nothing official is named Workspace, though Nous's Cloud page does use the word in passing for an agent's persistent file area.
What the project gives you is a cockpit for an agent you already run:
  • Chat with live streaming responses
  • A real terminal in the browser
  • A memory inspector you can read and edit
  • Skill browsing and management
  • Conductor and Swarm multi-agent modes
  • A Kanban task board for agent work, from backlog to done
  • Installable as a PWA on desktop and phone, with themes
Version 2 is "zero-fork": it runs against vanilla NousResearch/hermes-agent installed the official way. Per its README, you point it at the agent gateway (port 8642 by default) and the official dashboard API on 127.0.0.1:9119, and the Workspace UI itself serves on port 3000. Three ports, three jobs: 8642 is the agent, 9119 is Nous's admin panel, 3000 is the community cockpit.
One thing to know before you commit: this is still a small project. Its homepage, hermes-workspace.com, went dark for a stretch in September 2026 and is back online as the project's install-and-docs site. The code tells the same story: one primary maintainer, no human commits on the main branch since late August 2026, and v2.3.0 (May 2026) still the latest release, even as stars keep climbing. None of that makes it bad software, but it sits on top of a fast-moving agent, so pin what works and read the release notes before upgrading.

The three things people mean by "hermes workspace"

Search for hermes workspace and you will land on pages describing three different objects. Here is the map:
People say
They mean
Where it lives
Hermes Workspace, the product
The community web UI above
Port 3000, github.com/outsourc-e/hermes-workspace
The agent's workspace
The working directory where Hermes starts terminal commands and reads AGENTS.md
terminal.cwd in config, default: wherever you launched Hermes
The dashboard
Nous's built-in admin UI
hermes dashboard, opens 127.0.0.1:9119
If you just want to see what your agent is doing, you want the dashboard. If you want to drive the agent from a browser like a power tool, you want Hermes Workspace. And if you are debugging why the agent cannot find your AGENTS.md, you care about the working directory. The next section covers that one, because it is the least understood.

Hermes Workspace vs Hermes Desktop vs the other web UIs

The confusion does not stop at three meanings, because the ecosystem keeps shipping lookalikes. Sorting the names out as of October 2026:
  • Hermes Desktop is Nous's official desktop app for Windows, macOS, and Linux: the agent in a native window, no VPS or browser setup.
  • Hermes WebUI is a separate open-source browser chat UI, best known from Hostinger's one-click VPS template, which bundles it with the agent. It is a lighter app than the Workspace.
  • hermes-workspace-plugin, from the same developer as the Workspace, adds a tab to the official dashboard that embeds the chat of a Workspace you already have running (auto-detected on ports 3000-3002). A shortcut into the Workspace, not a replacement for running it.
  • hermes workbench is a community plugin added to the official plugin catalog in September 2026; it serves the Hermes Desktop app's UI in a browser. Also not a Nous product.
Rule of thumb: if you installed nothing extra, you have the dashboard. If you ran the installer from hermes-workspace.com, Docker Compose from the outsourc-e repo, or Hostinger's Hermes Workspace one-click template, you have the Workspace. Anything else is a different project wearing a similar name.

Where your Hermes agent's workspace actually lives on disk

The official docs define it plainly: "A workspace or working directory is where terminal commands start." It is controlled by terminal.cwd in ~/.hermes/config.yaml, and the default value is ., meaning the directory you launched Hermes from. One wrinkle: that key governs gateway and cron sessions, while the interactive CLI always starts where you launched it. There is no config key named workspace.
That directory is also where Hermes discovers your project context at session start. It loads one context type, first match wins: .hermes.md (or HERMES.md), then AGENTS.override.md, AGENTS.md, CLAUDE.md, .cursorrules. Inside a git repo, AGENTS.md files merge as a chain from the repo root down to your current folder, deeper files winning. Launch the agent from the wrong folder and it quietly runs without your instructions.
Here is the part that trips up people coming from OpenClaw: Hermes has no dedicated workspace folder. OpenClaw keeps AGENTS.md, MEMORY.md, and skills together in one workspace directory. Hermes splits them. The working directory holds your project and context files. Everything the agent owns lives in HERMES_HOME, which is ~/.hermes/ by default:
  • memories/ with MEMORY.md (the agent's notes, capped at 2,200 characters) and USER.md (its profile of you)
  • skills/ for installed skills
  • SOUL.md, the agent's persona, loaded from HERMES_HOME only
  • cron/ for scheduled jobs, plus state.db, the SQLite store that now holds every chat and messaging session (sessions/ lingers only as a legacy mirror)
Practical consequence: when you back up or migrate a Hermes box, back up ~/.hermes/, not just your project folder. The Docker backend adds one more layer: inside the container the workspace root is /workspace, and host folders only appear there if you mount them via docker_volumes or flip docker_mount_cwd_to_workspace to true, which mounts the folder you launched from. Hermes can also checkpoint the working directory before file changes and roll back with /rollback, but checkpoints are opt-in: start with --checkpoints or set checkpoints.enabled: true. Update backups are similar: updates.pre_update_backup defaults to a quick state snapshot, and only the full setting zips all of HERMES_HOME first.
For a deeper look at the container route, see our Hermes Docker guide, and for what skills actually are, the Hermes skills hub explainer.

How to run Hermes Workspace (the web UI)

The README documents three install paths:
  1. Docker Compose: clone the repo and bring the stack up alongside the agent.
  1. One-line installer: a script that installs both the agent and the Workspace on a fresh box.
  1. Attach mode: already running Hermes? Point the Workspace at your existing gateway and dashboard.
The README asks for Node 22+, an OpenAI-compatible backend, and Python 3.11+ only if the agent gateway runs on the same box; it lists no RAM number, though in practice you want 2 GB or more for the agent plus the UI. The quick starts claim one to three minutes; budget closer to half an hour the first time, once provider keys, a tunnel, and a fresh Ubuntu box are part of the job.
The security posture matters more than the install. Both the Workspace and the dashboard are designed for localhost. Since the June 2026 hardening, the official dashboard refuses non-loopback binds unless you configure a password or OAuth first, and the old --insecure flag no longer disables authentication. The Workspace ships auth middleware on every route and a fail-closed startup guard that refuses to bind beyond localhost until you set HERMES_PASSWORD, but that gate is far less battle-tested than the dashboard's, so the safe way to reach either one from your phone is a Tailscale network or an SSH tunnel, never a raw 0.0.0.0 bind on a public VPS.

What it costs to run in 2026

Hermes is free software and so is Hermes Workspace. What you pay for is model usage plus somewhere always-on to run them. The honest comparison:
Path
Price
What you actually get
Your laptop
$0
Agent dies when the lid closes; fine for trying it
Agent37 Free
$0/mo
One managed personal agent; sleeps after 15 min idle, wakes on a message
Hostinger one-click templates (Workspace or WebUI)
$6.49 to $25.99/mo promo
VPS templates deploying either the Workspace or the lighter WebUI + agent stack; renews at $11.99 to $49.99, ops are yours
Generic VPS
~$12 to $26/mo
Cheapest raw compute: a 2 GB box floors near $12, a 4 GB box is $20 to $26 at Hetzner, Vultr, or DigitalOcean (budget hosts go lower on long prepays); you do installs, updates, security, backups
Agent37 Basic
$3.99/mo
Managed always-on Hermes: 1 vCPU, 4 GB RAM, 8 GB disk, one-click deploy
To be clear about the DIY row: you do not need us. If you are comfortable with SSH, Docker, and being on call for your own server, a cheap VPS runs Hermes well, and our VPS-for-Hermes guide walks through exactly that. Managed hosting buys the absence of that job, not a capability you cannot replicate.
Nous also sells a first-party hosted option, Hermes Cloud: an always-on agent that scales to zero when idle (a stopped agent still bills $0.03/day for storage), in two sizes, Standard at $0.56/day and Pro at $1.09/day running, with a $2 credit minimum (or an active subscription) to deploy. Each Cloud agent gets its own persistent workspace, so files and memory survive stops. What it does not offer is your own UI: you reach it through the Portal's dashboard, the official Desktop app, or connected channels like Telegram and Discord, and there is no documented way to point the Workspace UI, or any custom frontend, at it. If the file browser and terminal are the point for you, you want a box whose ports you control.

Which path fits you

You want to monitor an agent you already run. Use the official dashboard. It is free, ships with the agent, and hermes dashboard opens it on 127.0.0.1:9119 with status, sessions, config, logs, and chat.
You want the full command center and enjoy tinkering. Self-host Hermes Workspace on your own hardware or VPS. Budget the setup time, keep it behind Tailscale, and pin versions.
You want the agent always-on without owning ops. Free to try with a personal plan (one agent that sleeps after 15 minutes idle), or deploy always-on managed Hermes at $3.99/mo. The instance comes with its own console: task board, full web terminal, visual file browser, and a live Linux desktop. The whole ~/.hermes state directory sits on a persistent volume, so memory, skills, and sessions survive restarts and updates. On the Cloud API side the platform documents exposing any instance port at its own HTTPS URL, so a self-run Workspace UI is possible there; on the managed plans the built-in console already covers most of what it does.
You are an agency running agents for clients. Give each client an isolated instance instead of stacking them on one shared VPS. The Agent37 Cloud API creates one per POST request, metered at $0.80 per vCPU, $0.70 per GB of RAM, and $0.09 per GB of disk per month: the standard 2 vCPU, 4 GB RAM, 4 GB disk shape comes to $4.76 a month always-on, white-label by default. More on that pattern in our Hermes agent API guide.

Frequently asked questions

Is Hermes Workspace an official Nous Research product?

No. It is an MIT-licensed community project by an independent developer (GitHub: outsourc-e). Nous Research's own interfaces are the built-in dashboard and the Hermes Desktop app, and its hosted product is Hermes Cloud. None of them is named Workspace.

Is hermes-workspace.com down?

Not anymore. The site went dark for a stretch in September 2026 but is back online as the project's official install-and-docs page. The GitHub repository remains the canonical home for code, issues, and releases.

What is the difference between Hermes Workspace and the hermes dashboard?

The dashboard is Nous's built-in admin panel on port 9119: status, sessions, config, logs. Hermes Workspace is a heavier community UI on port 3000 that adds streaming chat, a terminal, memory editing, and multi-agent orchestration. Both sit on top of the same agent.

Where does Hermes store memory and skills?

Not in the workspace. Memory (MEMORY.md, USER.md), skills, SOUL.md, sessions, and cron jobs all live in HERMES_HOME at ~/.hermes/. The workspace is just the directory where terminal commands start and context files load from.

Can I run Hermes Workspace on Hostinger?

Yes. Hostinger now offers two one-click Docker templates: a Hermes Workspace template that deploys the Workspace command center itself (its page lists the Conductor orchestrator, the browser terminal, and persistent memory plus skills), and a separate Hermes WebUI template that bundles the lighter chat UI with the agent. Both run on the same KVM plans, $6.49 to $25.99 per month at promo pricing with renewals roughly double, and either way you own updates, security, and backups.

What is the cheapest way to run all of this always-on?

Free, if an agent that sleeps between messages is acceptable: Agent37's Free tier runs one managed personal agent at $0/mo; it sleeps after 15 minutes idle and wakes when you message it. For always-on, a bare 2 GB VPS floors around $12 a month plus your time, a mainstream 4 GB box runs $20 to $26, and Agent37's Basic tier is $3.99/mo for a managed always-on Hermes with persistent storage, deployed in one click. Free software either way; what you are pricing is the box and who carries the pager for it.
Vishnu

Written by

Vishnu

Founder at Agent37, which runs managed hosting for OpenClaw and Hermes agents for 1,000+ users. Writes about what actually breaks when you leave an AI agent running.