Hermes Agent desktop and CLI upgrades, OpenClaw v2026.6.6 security fixes, and agent messaging tools
TL;DR
Hermes Agent and OpenClaw shipped dozens of agent interface, security, and messaging updates across the week that make local and multi-channel agents more reliable for daily use.
What shipped
Across this week Hermes Agent and OpenClaw released a steady stream of updates focused on agent interfaces, security boundaries, and messaging channels. The releases center on practical improvements to desktop apps, command tools, and provider integrations rather than new flagship models. Builders can now test these changes in existing workflows without waiting for larger platform shifts.
Hermes Agent
Hermes Agent delivered eleven targeted updates that improve daily operation of its autonomous agent framework. The changes cover interface choice, conversation control, onboarding, skills management, localization, and remote connectivity. Most updates address friction points reported by users running the agent across CLI, TUI, and desktop surfaces.
- •CLI vs TUI default Hermes Agent lets users set hermes chat to start in classic CLI or Ink TUI by default and supplies a --cli flag for one-off overrides in daily sessions.
- •Undo command The new /undo [N] command backs up N user turns, prefills the last message for edits, and removes intermediate turns in CLI, TUI, and messaging apps.
- •Fuzzy model search Model picker now uses fuzzy matching across desktop, web, TUI, and CLI while grouping providers and refreshing the catalog every hour.
- •Quick Setup flow First-time users sign in via Nous Portal for immediate chat access instead of stepping through the full configuration wizard.
- •NVIDIA skills tap NVIDIA/skills is now a default trusted source in the Skills Hub so CUDA-X, AIQ, and cuOpt skills appear in searches and update automatically.
- •Default skills cleanup Redundant built-in skills were removed and niche skills made optional with relevance gates to keep prompts shorter and more focused.
- •Simplified Chinese UI Hermes Desktop now ships with complete Simplified Chinese text in chat, sidebar, settings, and command center selectable in Appearance options.
- •Web Admin Panel The browser dashboard gained a full admin panel for MCP catalog, Telegram/Discord/Slack channels, credentials, and webhook configuration.
- •Remote gateway login Desktop app can connect to remote Hermes gateways over secure WebSockets using OAuth or username and password for multi-profile work.
- •Native desktop app A new native app for macOS, Linux, and Windows adds polished chat GUI, self-updates, drag-and-drop files, clipboard paste, and inline model picker.
OpenClaw
OpenClaw released thirty-nine updates dominated by security hardening, messaging reliability, and provider expansions. The bulk of changes tighten transcript and sandbox boundaries while adding support for new models and channels such as iMessage, Telegram, and QQBot. Multiple beta releases and one stable version rolled out during the week.
- •OpenRouter OAuth onboarding OpenClaw added OpenRouter OAuth sign-in and Claude Fable 5 adaptive thinking while keeping Gemma 4 reasoning replay intact.
- •iMessage recovery fixes Always-on inbound restart, durable echo markers, and startup diagnostics now reduce delivery failures on iMessage channels.
- •Transcript security tighten Security rules now cover transcripts, sandbox binds, MCP stdio, and Codex HTTP with exec approvals failing closed on timeout.
- •Feishu context leak fix Prompt-preface leakage into Feishu replies was blocked and WebSocket payload handling plus symlink writes were hardened.
- •Kimi K2.7 Code support Kimi K2.7 Code model was added with fixes for native tool-call IDs and improved Mistral and Fireworks parameter handling.
- •UI and mobile stability Workspace collapsing, WebChat backscroll, sidebar picker, and stale iOS gateway reconnects were improved for smoother sessions.
- •Telegram and WhatsApp delivery Telegram now supports rich text, tables, and blockquotes while WhatsApp honors ACP bindings for reliable delivery.
- •GLM-5.2 and Claude Haiku 4.5 New model support for GLM-5.2 and Claude Haiku 4.5 plus normalized IDs for OpenRouter and Google Vertex were added.
- •Control UI latency drop Cached metadata and lazy slash-command loading cut startup and first-reply latency in the Control UI.
- •Browser and MCP options Existing-session CDP support, WebSocket validation, and Streamable HTTP loopback transport were added for browser agents.
- •Telegram routing safety Account-scoped topic routing, streamed text survival during tools, and unauthorized DM exclusion improved Telegram reliability.
- •v2026.6.6 release note OpenClaw v2026.6.6 shipped with the security boundary tightening described in the same week.
- •SQLite auth migration Auth profiles moved to SQLite for durability and official npm plugins now keep trusted pins with fallback integrity checks.
- •Parallel search bundle Parallel was added as a bundled web search provider with automatic key discovery and onboarding picker support.
- •QQBot tag stripping Raw thinking tags are now stripped from QQBot replies and rich MCP tool results are coerced to avoid API errors.
- •QQBot beta release Beta 2026.6.5 introduced QQBot tag stripping, rich MCP coercion, and Parallel search in one package.
- •Parallel and SQLite beta Beta release added Parallel search, SQLite auth migration, and Anthropic extended-thinking recovery fixes.
- •Security and messaging beta Beta 2026.6.6 tightened boundaries, improved Telegram and iMessage delivery, and lowered Control UI latency.
- •Matrix voice notes Matrix integration gained voice note preflight checks and preserved thread reads through relations pagination.
- •Mobile provider status Android now shows expiring provider states and theme selection while iOS keeps diagnostics and gateway rows reachable.
- •MCP result coercion Resource links, audio, and malformed images are now coerced at the boundary to prevent Anthropic 400 errors.
- •v2026.6.5 release note OpenClaw v2026.6.5 shipped with QQBot stripping and Parallel search integration.
- •v2026.6.6-beta.2 tag Beta 2026.6.6-beta.2 was published highlighting ongoing security work.
What this means for you
For Vibe Builders: You can now run Hermes Agent on a native desktop app with drag-and-drop files and choose between CLI or TUI without setup friction. OpenClaw adds reliable messaging on Telegram and iMessage plus simpler model switching so you can ship agent workflows faster without writing code. Test the new /undo command and Quick Setup path this week to cut daily friction.
For Non-techies: For your business, Hermes Agent desktop and OpenClaw messaging updates mean agents can now handle chats on Slack, Telegram, and iMessage with fewer dropped messages. Security fixes reduce risk when the agent accesses tools or files. Sign in once via Nous Portal or OpenRouter OAuth and start using the same agent across phone and computer.
For Developers: On the platform side, Hermes Agent Web Admin Panel and OpenClaw SQLite auth migration plus tightened sandbox rules give clearer production controls. Evaluate the new remote gateway OAuth flow and Parallel search bundle against your current MCP and browser setups. Watch the next beta tags for continued iMessage and Matrix stability signals before rolling into production pipelines.
What to watch next
Watch for OpenClaw v2026.6.7 stable and any new Hermes Agent skills taps. Check whether Parallel search or Kimi K2.7 Code appears in production benchmarks this week.
Harsh’s take
The week shows both projects spending most effort on plumbing rather than new capabilities. Hermes Agent polished onboarding and desktop access while OpenClaw iterated on security and channel reliability across many small releases. The pattern suggests teams are still fixing basic agent reliability before broader adoption. Builders should install the native Hermes desktop app and run one OpenClaw beta on a non-critical channel to measure real stability gains instead of waiting for marketing announcements.
by Harsh Desai
Sources
- •[v2026.6.8-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.6.8-beta.1)
- •[v2026.6.7-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.6.7-beta.1)
- •[v2026.6.6-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.6.6-beta.2)
- •[v2026.6.6-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.6.6-beta.1)
- •[v2026.6.5-beta.6](https://github.com/openclaw/openclaw/releases/tag/v2026.6.5-beta.6)
- •[v2026.6.5-beta.5](https://github.com/openclaw/openclaw/releases/tag/v2026.6.5-beta.5)
- •[v2026.6.5-beta.3](https://github.com/openclaw/openclaw/releases/tag/v2026.6.5-beta.3)
More AI news
- Daily RoundupFERC fast-tracks AI data centers, Grok video models on Replicate, and agent builders for production
Infrastructure rules, new video generation models, and no-code agent platforms moved from announcement to usable tools on 18 June, shifting focus from chat interfaces to deployed systems.
- Model ReleaseAnthropic brings Artifacts to Claude Code for sharing live coding pages
Claude Code now converts session results into interactive artifacts that teams can share. The pages draw from full context, update automatically, and maintain version history.
- FeaturePerplexity releases Agent API to simplify multi-step AI agent workflows
Perplexity introduced the Agent API, a fully managed runtime designed to simplify the deployment and execution of complex multi-step AI agent workflows.