Hermes Agent v0.19.1 with GPT-5.6 and 80% latency cut, OpenClaw MCP apps (agent routing and recovery tools)

What shipped

From 27 July to 3 August 2026 both Hermes Agent and OpenClaw shipped repeated updates focused on reliability, model breadth, and daily usability. Hermes Agent concentrated on gateway routing, crash recovery, and desktop speed while OpenClaw emphasized meeting integration, Wear OS controls, and conversation branching. The combined releases give builders concrete ways to run agents across channels without losing messages or secrets.

Hermes Agent

Hermes Agent delivered the bulk of changes with fifty updates centered on v0.19.1. The release added GPT-5.6, Grok 4.5, and Claude Sonnet 5 through Fireworks AI and DeepInfra, introduced max and ultra reasoning tiers, and cut first-turn latency from 4.3 s to 0.9 s. Durable delivery ledgers, Bitwarden and 1Password integration, and smart approvals reduce message loss and approval fatigue for users running agents on Discord, Slack, and Telegram.

• •Model support expansion Hermes Agent added GPT-5.6, Grok 4.5, Claude Sonnet 5, and Kimi K3 via Fireworks AI and DeepInfra with max and ultra reasoning tiers for per-task control.
• •Gateway profile routing A single gateway now routes specific guilds or channels to isolated profiles that keep separate memory, skills, and secrets.
• •Durable delivery ledger Final responses are written to state.db before sending so unsent messages are redelivered after any gateway reboot.
• •Live subagent transcripts Task delegation now writes real-time transcript files that can be tailed and persists background results across restarts.
• •Bitwarden and 1Password integration API keys load from vaults using op:// references at startup, removing plaintext .env files.
• •Smart LLM approvals Flagged commands are reviewed by an independent LLM by default with custom deny rules that work even in yolo mode.
• •Subscription commands Users check and change Nous plans with /subscription and /topup directly in the terminal or desktop settings.
• •Desktop performance overhaul Twenty optimizations delivered 14x faster markdown streaming and virtualized diffs that prevent freezes on long transcripts.
• •Cold-start latency reduction First-turn time dropped 80 percent from 4.3 s to 0.9 s across CLI, TUI, gateway, and desktop with live reasoning streams.
• •v0.19.1 patch release Over 1,000 merged changes since v0.19.0 are now in a stable tagged build that fixes gateway, voice, and installer issues.
• •Session export tool The hermes sessions export command writes Markdown, HTML, Quarto, or Hugging Face traces with secret redaction and lineage stitching.
• •Provider catalog update Fireworks AI and DeepInfra joined the catalog alongside GPT-5.6, Grok 4.5, and local LM Studio JIT loading.
• •Crash-proof message delivery Responses are recorded in SQLite before transmission and resent automatically after any gateway restart.
• •Background task recovery Subagent dispatches write live logs and store completion results in a durable ledger that survives process restarts.
• •Secret source interface Pluggable SecretSource lets Hermes pull credentials from multiple Bitwarden or 1Password vaults with conflict warnings.
• •Default smart approvals An LLM reviewer evaluates flagged commands automatically and accepts user-defined deny rules that persist across sessions.
• •In-terminal plan management /subscription and /topup commands let users preview costs and schedule changes without leaving the CLI or TUI.
• •Virtualized desktop diffs Review panes now use virtual lists so large transcripts no longer cause layout thrashing during session switches.
• •Per-token response rendering Reasoning models stream thinking live and response boxes update token by token for faster perceived speed.
• •Stable v0.19.1 tag The patch release bundles fixes for gateway, voice, desktop, and installers after more than 1,000 pull requests.
• •Version 2026.7.30 tag Hermes Agent v0.19.1 (v2026.7.30) is now the recommended build for downstream deployments.
• •Reasoning effort overrides Max and ultra tiers can be pinned per model or per slot with task-specific auxiliary effort settings.
• •New model catalog entries GPT-5.6 variants, Grok 4.5, Moonshot Kimi K3, and Claude Fable 5 are now selectable in CLI and desktop.
• •Multi-tenant gateway routing One bot token can direct different channels to distinct profiles without misconfiguration crashes.
• •SQLite delivery ledger Responses are logged before platform send and redelivered on boot for Telegram, Discord, and Slack.
• •Live subagent monitoring Dispatch transcripts are written to files that survive restarts and can be tailed in real time.
• •Vault secret loading API keys are fetched at load time from Bitwarden or 1Password with deterministic precedence rules.
• •LLM-backed command review Flagged actions are checked by a separate reviewer and blocked by custom deny rules even in yolo mode.
• •Direct subscription controls Users view allowances and apply upgrades or downgrades inside the terminal without a browser.
• •Markdown streaming speed Lexing CPU usage fell 14x while virtualized diffs keep the desktop responsive on long sessions.
• •First-token latency cut Cold starts now finish in 0.9 s instead of 4.3 s with live reasoning and per-token painting enabled.
• •Multi-format session export Chat histories can be saved as Markdown, Quarto, HTML, prompt-only, or Hugging Face traces with redaction.
• •Expanded provider list Fireworks AI, DeepInfra, and GPT-5.6 variants are now first-class options alongside local JIT loading.
• •Isolated profile routing Specific channels or threads are sent to dedicated profiles that keep separate configurations and secrets.
• •Delivery obligation ledger Completed responses are stored before transmission and resent automatically after any crash.
• •Real-time subagent logs Task outputs stream to human-readable files that persist across process restarts.
• •Pluggable secret sources Credentials load from Bitwarden or 1Password vaults at startup without plaintext storage.
• •Automatic smart approvals An LLM reviewer handles flagged commands by default and relays refusal reasons back to the agent.
• •Terminal subscription tools /subscription and /topup commands show usage and let users change plans without leaving the keyboard.
• •Desktop rendering fixes Twenty performance PRs reduced re-renders and added pre-warmed backends for faster session switches.
• •Accelerated first-token times Response latency fell 80 percent across all platforms with live reasoning streams turned on by default.
• •Archive export command Session histories export to multiple formats with workspace filters and secret scrubbing included.
• •Catalog expansion GPT-5.6, Grok 4.5, Claude Sonnet 5, and Kimi K3 are now supported through new providers.
• •Channel-to-profile routing A multiplexed gateway directs messages to isolated profiles with separate memory and skills.
• •Durable gateway ledger Responses are recorded before send and redelivered on reboot to prevent silent loss.
• •Subagent transcript files Dispatches generate live logs that can be tailed and survive background restarts.
• •Bitwarden secret fetch API keys are pulled from vaults at load time via the SecretSource interface.
• •Default LLM approvals Flagged commands are reviewed automatically with custom deny rules that work in yolo mode.
• •In-app plan changes Users manage Nous subscriptions directly in CLI and TUI with cost previews.
• •Performance PR bundle Over twenty changes delivered 14x faster markdown and zero layout thrashing on session switches.

OpenClaw

OpenClaw shipped twenty-three updates that added MCP Apps, structured questions, Wear OS controls, and local inference. The changes focus on meeting plugins, durable delivery, and conversation branching so agents can join video calls and recover from crashes. Builders gain hosted Fish Audio synthesis and Claude Opus 5 support alongside crash-recoverable SQLite snapshots.

• •Fish Audio S2.1 integration Hosted synthesis now supports streaming voice notes, voice discovery, and telephony plus local S2 Pro reference streaming on macOS.
• •Claude Opus 5 and Kimi K3 support The runtime catalog added Claude Opus 5, Kimi K3, and GPT Live realtime feeds through Platform API paths.
• •Local llama.cpp setup Guided onboarding detects local providers and offers RAM-gated GGUF inference plus direct model downloads on Linux and macOS.
• •Wear OS companion app A phone-proxied watch app provides agent selection, realtime Talk controls, and an instant-talk tile on smartwatches.
• •Meeting plugin support Agents join Teams, Zoom, and Google Meet calls with default transcript collection and OpenAI or Gemini video feeds.
• •Structured agent questions Agents can now ask option-card questions across web, channels, and native apps with push-notification approvals.
• •MCP App hosting Ticketed Model Context Protocol apps can be pinned to durable dashboards and opened from channel replies.
• •Session rewind and branching Users can fork conversations from any message, switch branches, and restore prompt images after a fork.
• •Durable channel delivery Accepted messages survive gateway restarts via shared ingress drain and dead-letter recovery on multiple platforms.
• •SQLite snapshot recovery Persisted data is protected by a quarantine store and crash-recoverable snapshots that reject schema-upgrade data loss.
• •Quick Chat for macOS and Linux Streaming, dictation, and model controls are now available with signed updates and desktop integration.
• •Provider catalog expansion Claude Opus 5, Kimi K3, and GPT Live are now selectable through Codex OAuth authentication.
• •Guided local inference Browser, Linux, and macOS setups detect providers and suggest downloadable models with llama.cpp paths.
• •Wear OS session controls The companion app offers home-screen agent selection and audio-reactive playback on smartwatches.
• •Video call plugins Default meeting plugins collect transcripts while realtime Talk adds OpenAI and Gemini video support.
• •Structured questions rollout Option-card questions and enhanced approval history are now active across all platforms.
• •Ticketed MCP Apps Interactive apps with bound tools can be pinned to dashboards and hardened inside a shared sandbox.
• •Conversation fork support Users rewind or branch sessions across web and native apps while preserving queued sends.
• •Crash recovery for channels Messages remain recoverable across restarts on Telegram, Signal, Slack, and nine other platforms.
• •Quarantine store protection Persisted data survives database damage with rollback-writer snapshot recovery after schema upgrades.
• •v2026.7.2-beta.7 release The beta tag marks the latest build containing the meeting and MCP App features.
• •v2026.7.2-beta.6 release This beta introduced Wear OS controls and structured question support.
• •v2026.7.2-beta.5 release The beta added durable delivery and local inference setup improvements.

What this means for you

For Vibe Builders: You can now route specific channels to isolated agent profiles and recover messages after crashes without writing code. Hermes Agent v0.19.1 gives you 80 percent faster first responses plus Bitwarden secret loading so you stop managing .env files. OpenClaw MCP Apps and Wear OS controls let you pin dashboards and run agents from a watch while structured questions reduce manual approvals.

For Non-techies: For daily business use, agents now join Zoom or Teams calls automatically and keep working after a restart. Hermes Agent subscription tools and smart approvals cut the need to approve every command while OpenClaw local inference and Wear OS app let you control sessions from a phone or watch without extra setup.

For Developers: Production stacks gain durable SQLite ledgers, pluggable SecretSource interfaces, and crash-recoverable snapshots that survive gateway restarts. Hermes Agent 0.19.1 and OpenClaw both expose per-model reasoning overrides and MCP App hosting so you can benchmark latency drops and integrate local llama.cpp paths before rolling out to multiple messaging platforms.

What to watch next

Track the next Hermes Agent patch for additional model overrides and watch OpenClaw beta.8 for expanded video meeting plugins. Monitor whether MCP App dashboards gain native plugin declarations and whether both projects publish reliability benchmarks on the new delivery ledgers.

Sources

• •[v2026.7.2-beta.7](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.7)
• •[v2026.7.2-beta.6](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.6)
• •[v2026.7.2-beta.5](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.5)