Jonathon's AI Wiki

Hermes Desktop — Official Nous Research App (macOS / Windows / Linux)

9 min read

Source: hermes-desktop-official-nousresearch-2026-06-02.md — official product page hermes-agent.nousresearch.com/desktop, fetched 2026-06-02. Vendor: Nous Research (first-party). License: MIT (open source). Version: Hermes Agent v0.15.2. Platforms: macOS 12+, Windows 10/11, Linux (any distro). Refreshed 2026-06-02 with a RoboNuggets hands-on walkthrough (raw/Hermes_JUST_Dropped_their_Official_Desktop_App_No_Code_Needed.md, YouTube) and Nous’s launch post (raw/x-bookmarks-recent-digest-2026-06-03.md): the app was first demoed at Jensen Huang’s NVIDIA GTC keynote and is now in public preview.

Hermes Desktop is Nous Research’s official desktop application for Hermes Agent — a one-install, cross-platform front-end (macOS / Windows / Linux) that packages the full stateful-agent stack behind a native app: multi-channel presence, persistent memory, natural-language scheduling, subagent delegation, web/vision/TTS, and sandboxed execution. It is open source (MIT), ships at Hermes Agent v0.15.2, and routes paid usage through the new Nous Portal credit tiers. This appears to be the first first-party desktop surface from Nous ^[inferred — the hermes-desktop-dodo-reach article’s 2026-05-28 Open Question explicitly asked whether Nous would ship an official desktop; this is that answer] — distinct from the community clients that previously filled the gap, one of which (dodo-reach) confusingly shares the “Hermes Desktop” name.

Key Takeaways

  • Official + open source. First-party Nous Research app, MIT-licensed, at Hermes Agent v0.15.2. This is the platform vendor shipping its own GUI — a different category from the third-party front-ends the wiki already tracks (dodo-reach SSH client, Herm-TUI, Hermes Console).
  • Cross-platform, one install: macOS 12+, Windows 10/11, Linux (any distro). Broader reach than the macOS-only dodo-reach client, and a lower macOS floor (12+ vs that client’s 14+).
  • Six capability pillars frame the app — Connect / Remember / Schedule / Delegate / Search / Experiment — i.e., the whole Hermes stack surfaced in a GUI rather than the CLI.
  • Lives everywhere: one agent + one memory across Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI (a growing list).
  • Persistent memory that learns your projects, auto-generates skills, and “never forgets how it solved a problem” — the GUI face of Hermes’s three-tier memory model.
  • Subagent delegation: isolated subagents with their own conversations, terminals, and Python RPC scripts for “zero-context-cost pipelines” — the desktop expression of the v0.15 Kanban-swarm delegation pattern. ^[inferred cross-mapping]
  • Sandboxing across five backends: local, Docker, SSH, Singularity, Modal — with container hardening + namespace isolation (the seven-layer security model exposed as a product feature).
  • Nous Portal monetization: Free / Plus / Super / Ultra; paid tiers bundle monthly credits usable across Hermes Agent + Nous Chat, “access to hundreds of cutting-edge models,” and higher usage limits.

The six pillars (page framing)

  1. Connect — Lives Everywhere. Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI + growing. “One agent, one memory, every surface.” (See the channel use-cases in Hermes User Stories.)
  2. Remember — Persistent Memory. Learns your projects, auto-generates skills, never forgets how it solved a problem.
  3. Schedule — Focused Automation. Natural-language scheduling for reports, backups, and briefings, running unattended through the gateway.
  4. Delegate — Tasks Multiplied. Isolated subagents with their own conversations, terminals, and Python RPC scripts for zero-context-cost pipelines.
  5. Search — Browse the Web. Web search, browser automation, vision, image generation, text-to-speech, and multi-model reasoning.
  6. Experiment — Isolated Sandboxing. Five backends (local, Docker, SSH, Singularity, Modal) with container hardening and namespace isolation.

Platforms & getting it

  • macOS 12+ · Windows 10/11 · Linux (any distro) — single app, all three.
  • Open source, MIT · current version Hermes Agent v0.15.2 (the desktop app version tracks the agent release line, suggesting a unified codebase + GUI rather than a separate product ^[inferred]).
  • The product page (hermes-agent.nousresearch.com/desktop) is the entry point; exact download/auto-update and code-signing/notarization mechanics are not stated in the captured page copy (see Open Questions).

Nous Portal tiers

  • Free / Plus / Super / Ultra.
  • Paid tiers include monthly credits for use in Hermes Agent & Nous Chat, access to hundreds of cutting-edge models, and high usage limits. Per-tier credit amounts and pricing are not listed on this page.

Hands-on walkthrough & launch context

  • Launch framing (Nous): “The next evolution of Hermes Agent… everything you love about Hermes, now native on your machine.” It was first demoed at Jensen Huang’s NVIDIA GTC keynote and is now in public preview. ^[from @NousResearch launch post, x.com/NousResearch/status/2061843507417944552 — ties to the NVIDIA-Nous partnership cluster in velocity-release]
  • Sub-5-minute, no-terminal setup (RoboNuggets walkthrough): download the installer for your OS from the product page, run it, and Hermes is set up “without you having to interface with the terminal” — the on-ramp the CLI/VPS route never had.
  • Familiar GUI for anyone who’s used the Codex or Claude Code desktop apps: start a chat session, browse default/installed skills, and a toolsets view.
  • Batteries-included skills out of the box — several dozen pre-installed, including a humanizer (removes AI-tells from writing), Excalidraw (diagrams), and connectors for Google Workspace, Linear, Notion, and PowerPoint. A strong default for beginners who’d otherwise install skills by hand.
  • Toolsets surfaced in the UI: cron jobs (scheduled background runs), code execution, web browsing, and ask-clarifying-questions — configurable or left at defaults.
  • GUI messaging setup, no CLI: wire Telegram / WhatsApp / etc. one-by-one from a Messaging pane. The Telegram flow is entirely in-app — paste a BotFather bot token and your allowed user IDs (so the agent only answers you), enable, save. No config.yaml editing.
  • Positioning: the walkthrough frames it as a Codex / Claude Code alternative for people who want a 24/7 messaging-native agent without terminal setup — the GUI expression of the “lives everywhere” pillar.

Operator walkthrough & remote-gateway tuning (2026-06-07)

A comprehensive third-party walkthrough — Alex Finn on Greg Isenberg’s channel (raw/Hermes_Agent_Desktop_-_Full_Setup_+_Real_Use_Cases.md), with corroborating operator posts in raw/x-bookmarks-recent-digest-2026-06-07.md — frames the v0.16-line desktop app as “the single best way to use Hermes Agent,” explicitly over the Telegram/Signal messaging surfaces (and, in the presenter’s framing, the moment Hermes “overtook OpenClaw”). Two concrete additions beyond the launch coverage above:

  • Remote-gateway architecture (the “control surface” pattern). Run the actual agent on your main machine / homelab / server (with all projects, skills, and memories) and connect the desktop app as a thin client from a laptop — they link over a secure WebSocket with OAuth, and each profile can point to its own remote host, so your API keys + compute stay where you want them. Setup is ~5 minutes (@tonbistudio, @HermesAgentTips). This is the architecture behind the v0.16 “remote-gateway connect” line item.
  • Operator tuning tips (@AlexFinn, “7 things”): run Hermes on your main computer (not a side machine) so it works alongside your real files; use /background <prompt> to fire concurrent background tasks; create a profile per model (e.g. Opus for writing/research, GPT for coding, a local model for private work) since each profile carries its own memories/skills/tools; plug in local models (e.g. Qwen on a Mac Studio / DGX Spark) via a dedicated profile; prune stale cron jobs (too many slow the agent); and lower the memory compression threshold (e.g. to 0.5) to compress more often and reduce memory loss on long projects.
  • Desktop-vs-terminal token trade-off (2026-06-09 creator guides). [YouTube signal — 2026-06-09] Two independent v0.16 walkthroughs (raw/Hermes_Agent_Update_v0.16_is_HUGE_Surface_Release.md, raw/The_Complete_Guide_to_the_Hermes_Agent_Desktop_App.md) flag a concrete cost criterion: the desktop app bleeds tokens on heavy UI schemas / backgrounds, while terminal/CLI use passes raw, highly-filtered text streams to the model and is more token-efficient. Operator rule of thumb: use the GUI for everyday drag-and-drop config / skills / session management; drop to the terminal for heavy coding or high-volume work where token efficiency matters.
  • Ollama launch path (2026-06-09). [X signal] Nous added ollama launch hermes-desktop — launch the desktop app directly through the Ollama local-runtime workflow, lowering local-model setup friction beyond the per-profile Qwen/DGX route above; the announcement ties it to Hermes’s self-learning Python skills (generate/improve skills from natural language). (Source: raw/x-account-nousresearch-2064468385748951415.md.)

Why It Matters

  • Nous goes first-party on the desktop. Until now the polished desktop/UI experience came from the community (dodo-reach, Herm-TUI, Console). An official cross-platform app dramatically lowers the on-ramp versus the CLI/VPS route taught in the 1-hour course — install an app instead of provisioning a server.
  • It answers an open question. The dodo-reach article (2026-05-28) flagged “worth watching whether Nous Research ships an official desktop surface.” It now has — and the overlap (sessions, cron, skills, channels) the community tools cover is now also first-party.
  • Naming collision is real. Two distinct things are now called “Hermes Desktop”: this official Nous app and dodo-reach’s third-party macOS SSH client. They are not the same product — disambiguate when referencing either.
  • Nous Portal is the monetization layer. Free→Ultra managed credits spanning Hermes Agent + Nous Chat is how Nous packages paid access to its hosted models around the open-source agent — the same Portal surfaced in the v0.15 velocity release.

Try It

  1. Download from hermes-agent.nousresearch.com/desktop for your OS (macOS 12+, Windows 10/11, or Linux).
  2. Connect a channel — wire Telegram / Discord / Slack / WhatsApp / Signal / Email so one agent + one memory follows you across surfaces.
  3. Start on the Free tier; upgrade through Nous Portal (Plus/Super/Ultra) when you need monthly credits, more models, or higher limits.
  4. Use the sandboxing (local → Docker → SSH → Singularity → Modal) to run experiments without risking the host — the GUI exposure of the security model.
  5. Schedule a recurring report/briefing in natural language to exercise the unattended-automation pillar.
  6. If you already self-host over SSH, compare against dodo-reach’s SSH client / Herm-TUI — the official app is the managed on-ramp, the community clients are the power-user host-direct surfaces.

Open Questions

  • Download / signing / auto-update mechanics. The page doesn’t state installer format, code-signing/notarization status (cf. dodo-reach’s ad-hoc-signed caveat), or how updates ship.
  • Per-tier pricing. Nous Portal lists Free/Plus/Super/Ultra but no credit amounts or prices on this page.
  • Codebase relationship. Whether the desktop app is the existing CLI Hermes Agent + a GUI shell (the v0.15.2 version match suggests unified ^[inferred]) or a distinct build is unconfirmed.
  • Effect on community clients. Does the official app supersede, coexist with, or complement dodo-reach / Herm-TUI / Console? Likely coexist (different audiences) but unstated.
  • Exact launch date. Resolved (2026-06-02): the app was first demoed at Jensen Huang’s NVIDIA GTC keynote and is now in public preview (per Nous’s launch post). It ships on the v0.15.x line; the precise public-preview calendar date isn’t stated.

Graph View

Backlinks