Source: ai-research/hermes-v0-16-0-surface-release-changelog-2026-06-05.md (official GitHub release notes, github.com/NousResearch/hermes-agent/releases/tag/v2026.6.5); announcement + Tool Search demo via raw/x-account-nousresearch-2063075092859637888.md and raw/x-account-nousresearch-2060490936124105179.md (@NousResearch). Unlike the v0.15 article’s creator-reported figures, the numbers here are from the official changelog.
Hermes Agent v0.16.0 (“The Surface Release”, 2026-06-05) is a major release — 874 commits, 542 merged PRs, 1,962 files changed, 399 issues closed (2 P0 / 62 P1 / 16 security-tagged), and 170 community contributors since v0.15.2. The name states the theme: Hermes grows real surfaces beyond the terminal — a native desktop app on macOS/Linux/Windows, a feature-complete web dashboard admin panel, and broader messaging reach — turning “the black screen of weird code” into installable, self-managing software. Install/update via hermes update; desktop at hermes-agent.nousresearch.com/desktop.
Key Takeaways
- Native desktop app (macOS / Linux / Windows) is the headline. One-click install with in-app self-updates, drag-and-drop files, clipboard image paste, streaming chat, a Cmd+K command palette, session list with archive/search, an inline status-bar model picker, full Simplified Chinese i18n, and remote-gateway connect (OAuth or username/password) so the desktop client can drive a remote Hermes instance. Multi-profile concurrent sessions with cross-profile
@sessionlinks. This is the “surface” that retires the terminal-only experience. - Web Dashboard becomes a full admin panel. A complete browser management plane: MCP catalog enable/disable toggles, a Channels page to configure every gateway messaging platform, credential management, webhook/hook creation, memory config, system settings, a check-before-update flow + Debug Share, bulk session ops, and a schedule picker. (Graduates the v0.15 “Web Dashboard overhaul” increment to feature-complete.)
- “Tool Search” = progressive tool disclosure. MCP and plugin tools are now disclosed progressively — agents dynamically load only the tools they need instead of front-loading every tool into context, cutting per-turn overhead. Same dynamic-tool-loading pattern as Claude’s own deferred-tool / ToolSearch mechanism.
- Quick Setup via Nous Portal — an install-to-first-message-in-seconds onboarding path (
hermes portal), plus a traditional Full Setup wizard, a “choose provider later” skip, and local/custom-endpoint config with no API key required. /undo [N]across every surface — backs up N user turns (prefill + soft-delete) in CLI, TUI, and messaging platforms. Plus/titlefor session renaming and/debugnow bundlingdesktop.log.- Skills system overhaul. A leaner default skill set; a new
environments:frontmatter relevance gate (kanban / docker / s6) that keeps context-specific skills out of the index until relevant; usage tracking so the curator can prune unused built-ins;install --no-skillsfor a blank slate; NVIDIA/skills is now a built-in trusted Skills Hub tap alongside OpenAI, Anthropic, and HuggingFace; new optional skillsgrok(xAI Grok Build CLI) andantigravity-cli. - New models + faster catalog.
deepseek-v4-flash,MiniMax-M3(1M context),qwen3.7-plus,gemini-3.5-flash; a fuzzy-search model picker on every surface; hourly catalog refresh (was daily); mid-session model switching persists to the DB with post-interrupt recovery. - Kanban / multi-agent matures.
goal_modecards run workers in a/goalloop (graduates the v0.15 increment), file/image attachments on tasks feed worker vision, aPOST /runs/{run_id}/terminateendpoint, and uncapped delegation depth (max_spawn_depthfloor 1, no ceiling). - Auth grows up: pluggable username/password login, generic self-hosted OIDC provider support,
hermes dashboard registerfor self-hosted OAuth clients, and xAI Grok OAuth as a first-class launcher provider. - Security: a CVE patch + 16 security-tagged fixes. CVE-2026-48710 (Starlette BadHost) pinned to Starlette ≥1.0.1; URL SSRF checks moved off the event loop; docker restart/stop/kill added to
DANGEROUS_PATTERNS; invisible-unicode sanitization in vetted skill content; Bedrock bearer token stripped from subprocess env. Extends the Hermes Security Model. - Perf: ~14% fewer tokens per
read_filevia a compact line-number gutter — the official confirmation of the @Teknium creator-reported figure in the v0.15 increments — plushermes sessions optimize(FTS5 segment merging) and/branchsessions staying visible after a parent reopen. - Post-release surface — iMessage via Photon (2026-06-08). [X signal] Extending the v0.16 remote-gateway, Nous added an iMessage interface: run
hermes gateway setup, select Photon as the gateway provider, and the agent answers in iMessage — a lightweight texting surface on top of the existing desktop / web / Telegram / Signal reach. (Source:raw/x-account-nousresearch-2064102412076364207.md.)
Notable verifications of prior v0.15 increments
The official v0.16 changelog confirms three claims that the v0.15 article could only mark as creator-reported ^[inferred]:
- ~14% read-file token savings (was @Teknium, 2026-05-30) → official: “compact line-number gutter in read_file (~14% fewer tokens per read).”
- Web Dashboard as a full management plane (was the 2026-06-03 “overhaul” increment) → official: the feature-complete admin panel above.
- Kanban jobs driven by
/goal(was @Teknium, 2026-05-31) → official:goal_modecards run workers in the/goalloop.
Why it matters
Where v0.15 was a maintainability-and-speed release (the 16K→3.8K-line refactor), v0.16 is a reach release — it puts Hermes in front of non-terminal users. A native desktop app with self-update, a browser admin panel that removes the need to run CLI commands, OIDC + username/password auth, and a seconds-long Nous Portal setup together lower the install-and-operate barrier that kept self-hosted agents in the hands of CLI power users. The scale (874 commits, 170 contributors in one cycle) is itself the signal: this is a fast-moving OSS project whose contributor base now rivals its maintainers. For the practical-AI reader, the load-bearing changes are progressive tool disclosure (cheaper context per turn), the environments: skill gate (less index noise), and the desktop + dashboard surfaces (operate a fleet without the terminal).
Try It
hermes updateto v0.16 (or grab the desktop app athermes-agent.nousresearch.com/desktop); on first run try Quick Setup via Nous Portal for the seconds-to-first-message path.- Open the Web Dashboard and manage the MCP catalog, channels, and credentials from the browser instead of editing config files.
- Watch progressive tool disclosure in a session with many MCP/plugin tools — confirm tools load on demand rather than all up front.
- Drive a Kanban
goal_modecard and confirm workers run in the/goalloop with file/image attachments feeding worker vision. - If self-hosting for a team, wire username/password or self-hosted OIDC via
hermes dashboard register.
Related
- Hermes Agent v0.15 — The Velocity Release — the predecessor; v0.16 confirms several of its creator-reported figures.
- Hermes Desktop (official Nous app) — the native desktop surface this release ships across all three OSes.
- Hermes MCP Catalog — now managed with enable/disable toggles from the web dashboard.
- Hermes Skill Bundles — the skills system overhauled here (environments gate, usage tracking, NVIDIA tap).
- Hermes Security Model — the defense-in-depth model the CVE patch + 16 security fixes extend.
- Nous Portal — the subscription that powers the Quick Setup path.
- Set Up a Grok Subscription in Hermes Agent — xAI Grok OAuth, now a first-class launcher provider.
- NVIDIA NemoClaw — context for NVIDIA/skills becoming a built-in trusted Skills Hub tap.
Open Questions
- The changelog was captured via WebFetch structured-extraction of the GitHub release page; spot-check exact command syntax (
hermes portal,hermes dashboard register,hermes sessions optimize) againsthermes-agent.nousresearch.com/docsbefore quoting in production. - “Progressive tool disclosure” is assumed equivalent to the separately-demoed “Tool Search” (@NousResearch, early June) — both describe dynamic per-need tool loading; confirm they are the same subsystem and not two mechanisms.